multiplier slow converge

[hailinzeng]

Hi, Jon

I just found another interesting trace that multiplier is stuck at state normalize_2 for a long time.

input:
a : 32'h5e_c72e (denormal)
b : 32'h1f_ddeb (denormal)

I have uploaded the testbench in here.

[hailinzeng]

In this trace, z_e is 0x302 when state becomes normalize_2. $signed(z_e) is -254, which requires 128 cycles to change the state to round.

fpu/multiplier/multiplier.v

Lines 209 to 220 in 97a7682

	normalise_2:
	begin
	if ($signed(z_e) < -126) begin
	z_e <= z_e + 1;
	z_m <= z_m >> 1;
	guard <= z_m[0];
	round_bit <= guard;
	sticky <= sticky | round_bit;
	end else begin
	state <= round;
	end
	end

It may not be considered as a bug. But it seems to be a good target for optimization.

[dawsonjon]

Yes, I think that could be a good target for optimisation. So I will leave this issue open for now as a reminder. However if you are looking for some fast code... you could take a look at my other project which provides a fully pipelined multiplier with a 9 clock cycle latency.