Interpreter crash from `tf.io.decode_raw`

Impact

The implementation of tf.io.decode_raw produces incorrect results and crashes the Python interpreter when combining fixed_length and wider datatypes.

import tensorflow as tf

tf.io.decode_raw(tf.constant(["1","2","3","4"]), tf.uint16, fixed_length=4)

The implementation of the padded version is buggy due to a confusion about pointer arithmetic rules.

First, the code computes the width of each output element by dividing the fixed_length value to the size of the type argument:

int width = fixed_length / sizeof(T);

The fixed_length argument is also used to determine the size needed for the output tensor:

TensorShape out_shape = input.shape();
out_shape.AddDim(width);
Tensor* output_tensor = nullptr;
OP_REQUIRES_OK(context, context->allocate_output("output", out_shape, &output_tensor));

auto out = output_tensor->flat_inner_dims<T>();
T* out_data = out.data();
memset(out_data, 0, fixed_length * flat_in.size());

This is followed by reencoding code:

for (int64 i = 0; i < flat_in.size(); ++i) {
  const T* in_data = reinterpret_cast<const T*>(flat_in(i).data());

  if (flat_in(i).size() > fixed_length) {
    memcpy(out_data, in_data, fixed_length);
  } else {
    memcpy(out_data, in_data, flat_in(i).size());
  }
  out_data += fixed_length;
}