Yeah, I remember this part being a bit shady, my apologies 
To avoid an issue with the 16-bit bytecnt, when you issue a read command I believe it was designed to actually empty the entire FIFO. The use of the ‘stream’ flag effectively overrides any byte count, see line 345 of reg_main.v:
if ( (reg_stream == 1) || ((bytecnt < (total_bytes-16'd1)) && (total_bytes != 0)) ) begin
This is why the system FIRST reads the number of bytes waiting before actually reading them. This way it knows when to stop the USB transfer based on how many bytes should be dumped over the USB line.
But I don’t know if I ever tested it with FIFOs that large, so there might be an error causing some problem there. The largest I had used was with the SASEBO-W which I had a MAX_SAMPLES of 49K samples too. Looking at the code around line 354 in reg_main.v is suspicious:
//Will select next byte
bytecnt <= bytecnt + 16'd1;I bet if you add code that pegs it to 0xFFFF instead of rolling over, it might work? From memory I thought it DID peg at 0xFFFF, so that might have been an oversight by myself.
Regards,
-Colin
EDIT: PS I feel I should explain this entire module came from when I was using the OpenADC on a serially-connected FPGA. Unfortunately it added some craziness to the design (i.e. I wanted almost zero overhead, the protocol had to work over a serial line), which wasn’t helped by grafting this onto several different USB solutions! But it hasn’t fallen over badly enough to bother rewriting…