autobatch is 10 times slower than manual batching?

I am trying autobatch in mnist example as below:
#if 0 //Manual Mini-batching
for (unsigned idx = 0; idx < bsize; ++idx) {
cur_batch[idx] = input(cg, {N_INPUT}, mnist_train[id + idx]);
cur_labels[idx] = mnist_train_labels[id + idx];
}

// Reshape as batch (not very intuitive yet)
Expression x_batch = reshape(concatenate_cols(cur_batch), Dim({N_INPUT}, bsize));
Expression loss_expr = nn.get_nll(x_batch, cur_labels, cg); // Get negative log likelihood on batch
#else //Automatic Mini-batching
vector lossVec;
for (unsigned idx = 0; idx < bsize; ++idx) {
cur_batch[idx] = input(cg, {N_INPUT}, mnist_train[id + idx]);
cur_labels[idx] = mnist_train_labels[id + idx];
Expression lossExpr = nn.get_nll(cur_batch[idx], cur_labels[idx], cg);
lossVec.push_back(lossExpr);
}
Expression loss_expr = sum(lossVec);
#endif
cg.forward(loss_expr);
loss += as_scalar(loss_expr.value()); // Get scalar error for monitoring
num_samples += bsize;
cg.backward(loss_expr); // Compute gradient with backward pass
trainer.update();

I did autobatch with --dynet-mem 1024 --dynet-autobatch 1 and run it on CPU

But the autobatch is 10 times slower than manual batching!
What had I done wrong?

Thanks a lot!

Start with the MNIST example and compare the manual and automatic mini-batching paths shown in the issue on CPU using --dynet-mem 1024 --dynet-autobatch 1. Measure where the automatic path incurs its extra time and verify whether the reported slowdown is reproducible; done means identifying the cause and recording or applying the appropriate correction.

Written by the indexing model from the issue text.