Generating a desired pulse train in Xilinx ISE software

What you want to do is called a Numerically Controlled "Oscillator", or NCO. It works like this...

Create a counter that can increment by values other than 1. The inputs to this counter are the master clock, and a value to count by (din). For each clock edge, count <= count + din. The number of bits in din is the same as the number of bits in the counter. The actual count value can be used for many useful things, but what you want to do is super simple.

You want to detect every time the counter rolls over, and output a pulse to your motor when that happens. Do this by taking the most significant bit of the counter and running it through a single flip-flop to delay it by one clock. Now you have two signals that I'll call MSB, and MSB_Previous. You know if the counter has rolled over because MSB=0 and MSB_Prev=1. When that condition is true, send a pulse to the motor.

To set the pulse frequency, the formula is this: pulse_rate = main_clk_freq * inc_value/2^n_bits

Where inc_value is the value that the counter is being incremented by and n_bits is the number of bits in the counter.

An important thing to note is that adding bits to the counter does not change the range of the output frequency-- that is always 0 Hz to half of main_clk_freq. But it does change the accuracy that you can generate the desired frequency. Odds are high that you won't need 32-bits for this counter, and that maybe just 10 to 16 bits will be enough.

This method of generating pulses is nice because it is super easy, the logic is small and fast, and it can often generate frequencies more accurately and with better flexibility than the type of counter+comparator design that you have in your question.

The reason why the logic is smaller is not only because you can get by with a smaller counter, but you do not have to compare the entire output of the counter. You only need the top bit. Also, comparing two large numbers in an FPGA usually requires a lot of LUTs. Comparing two 32-bit numbers would require 21 4-Input LUTs and 3 logic levels, where as the NCO design requires 1 LUT, 2 Flip-Flops, and only 1 logic level. (I'm ignoring the counter, since it is basically the same for both designs.) The NCO approach is much smaller, much faster, much simpler, and yields better results.

Update: An alternative approach to making the rollover detector is to simply send out the MSB of the counter to the motor. If you do this, the signal going to the motor will always be a 50/50 duty cycle. Choosing the best approach depends on what kind of pulse your motor needs.

Update: Here is a VHDL code snippet for doing the NCO.

signal count :std_logic_vector (15 downto 0) := (others=>'0);
signal inc   :std_logic_vector (15 downto 0) := (others=>'0);
signal pulse :std_logic := '0';

. . .

process (clk)
begin
  if rising_edge(clk) then
    count <= count + inc;
  end if;
end process;

pulse <= count(count'high);

It sounds as though the drive from the optocoupler is marginal for the FPGA.

The fact that the scope probe loads it enough to make things much worse seems to confirm it. This is probably capacitance rather than resistance. For example, a Tek P2200 1x/10x passive probe in the 1x position can put 80 pF on the circuit being probed. At 100 kHz, that's -j 20K ohms, which MAY be OK. At 1 MHz (100 kHz x 10), that's -j 2K ohms, which probably isn't.

You really don't want to know why I can quote those numbers for that probe right now. I didn't take a hatchet to the dam' thing, but I was sorely tempted.

Take a look at the current drive specs for the optocouplers. Consider using buffers. If the optocouplers are open drain (open collector) devices, you NEED pull-up resistors, and the smaller the resistance you can get away with, the crisper your rising edges will be.