[luci-interpreter] ReshapeOp with 0 and -1 in shape

[seanshpark]

ReshapeOp from ONNX has 0 and from this value, luci-interpreter halts without message.
maybe by divide by zero error.

static void resolveUnknownDimension(const Shape &input_shape, Shape *output_shape)
{
  const int32_t num_input_elements = input_shape.num_elements();
  int32_t num_output_elements = 1;
  int unknown_dim_index = -1;
  for (int i = 0; i < output_shape->num_dims(); ++i)
  {
    const int32_t value = output_shape->dim(i);
    if (value == -1)
    {
      assert(unknown_dim_index == -1);
      unknown_dim_index = i;
    }
    else
    {
      num_output_elements *= value;
    }
  }
  if (unknown_dim_index != -1)
  {
    output_shape->dim(unknown_dim_index) = num_input_elements / num_output_elements; <-----------
    num_output_elements *= output_shape->dim(unknown_dim_index);
  }

what to do

• throw with there's a 0 in shape message
• treat 0 like the ONNX does

[seanshpark]

https://onnx.ai/onnx/operators/onnx__Reshape.html

allowzero - INT (default is '0'):
(Optional) By default, when any value in the ‘shape’ input is equal to zero the corresponding dimension value is copied from the input tensor dynamically. allowzero=1 indicates that if any value in the ‘shape’ input is set to zero, the zero value is honored, similar to NumPy.

[seanshpark]

luci shape inference

        else if (const_shape_node->at<S32>(axis) == 0)
        {
          const auto node_tensor = loco::must_cast<luci::CircleNode *>(node->tensor());
          // set dim value to input
          if (node_tensor->shape_status() == luci::ShapeStatus::VALID && axis < node_tensor->rank())
            shape_by_input.dim(axis) = node_tensor->dim(axis);
          else
          {
            // stop to check if this case exist for debugging
            INTERNAL_EXN("Check Reshape shape with 0");
          }
        }

[seanshpark]

For testing, I've copied like this.
But output_tensor after kernel.configure(); still has 0, which makes
output_tensor size 0 tensor.

TEST_F(ReshapeTest, ZeroDimension)
{
  Shape input_shape{2, 1, 2, 3};
  std::vector<float> input_data{1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12};
  Shape shape_shape{4};
  std::vector<int32_t> shape_data{0, 2, -1, 3};
  Tensor input_tensor =
    makeInputTensor<DataType::FLOAT32>(input_shape, input_data, _memory_manager.get());
  Tensor shape_tensor =
    makeInputTensor<DataType::S32>(shape_shape, shape_data, _memory_manager.get());
  Tensor output_tensor = makeOutputTensor(DataType::FLOAT32);

  Reshape kernel(&input_tensor, &shape_tensor, &output_tensor);
  kernel.configure();
  _memory_manager->allocate_memory(output_tensor);
  kernel.execute();

  EXPECT_THAT(extractTensorData<float>(output_tensor), FloatArrayNear(input_data));
}

[seanshpark]

change

--- a/compiler/luci-interpreter/src/kernels/Reshape.cpp
+++ b/compiler/luci-interpreter/src/kernels/Reshape.cpp
@@ -61,7 +61,11 @@ static void resolveUnknownDimension(const Shape &input_shape, Shape *output_shap
   int unknown_dim_index = -1;
   for (int i = 0; i < output_shape->num_dims(); ++i)
   {
-    const int32_t value = output_shape->dim(i);
+    int32_t value = output_shape->dim(i);
+    if (value == 0)
+    {
+      value = input_shape.dim(i);
+    }
     if (value == -1)
     {
       assert(unknown_dim_index == -1);

[seanshpark]

Change may take some time... in the meantime, throw some error messages, instead of silent exit.

[seanshpark]

what happens output_shape.size() > input_shape.size() ?