The QAM demodulation in matlab you used is possibly not correct.
It loss the system performance about 6~10dB Eb/No in the same average bit error rate & code rate condition.

The correct way is to tell the matlab how much is the variance of AWGN in the system.
You can change function qamdemod() In ldpcExample.m as follows:
dataLlr = qamdemod(dataRx, M, 'OutputType', 'approxllr', 'UnitAveragePower', true,'NoiseVariance',10^(-snr/10));

This helps to get the log-likelihood ratio more precisely.

On the other hand, the speed of the code run is too slow, we can call

h_dec = comm.gpu.LDPCDecoder(sparse(LDPC.H==1),'MaximumIterationCount',500);

and use
llr_vec = [zeros(1,2*LDPC.Z), dataLlr'];
dataHat = step(h_dec,llr_vec');
to simulate faster in the matlab.