module averager(

input  [9:0] n;
input clk;
input rst;
output reg [19:0] out;
output reg [9:0] cnt;
output reg [19:0] sum;
output reg [9:0] avg;

integer i = 0;

always @(posedge clk ) 
    if (rst == 1) begin 
        sum = 20'b0;
        cnt = 10'b0;
        out = 20'b0; 
        avg = 10'b0;
    end else if (rst == 0) begin
        sum = sum + n;
        out = sum;
        cnt = cnt + 1;
        avg = 0;

        for (i=0; i<641; i=i+1) begin
            if(out >= cnt) begin
                out = out - cnt;
                avg = avg + 1;

The above is the code to implement a cumulative moving average filter. The for loop is used for division to find the average and involves repeated subtraction. However I am getting the following warning and error:

WARNING:Xst:2254 - Area constraint could not be met for block <averager>, final ratio is 509.
WARNING:Xst:1336 -  (*) More than 100% of Device resources are used
ERROR:Pack:18 - The design is too large for the given device and package. 
  Please check the Design Summary section to see which resource requirement for
  your design exceeds the resources available in the device.

This must be because I am using large values in the for loop and thus am getting a large circuit that can't be implemented. I am looking for an alternative of the for loop, which could find the average for me. I just need the quotient value.

Design Properties: Family: Spartan3E Device: XC3S500E


You are correct with the guessing the for loop.

The for-loop logic is huge when it after it static unrolls. With your current code, you cannot handle the worst case scenario where n=1023. To cover this with your current code you'd need a for loop with 1024 iterations.

Instead of a up counter, you can use a down counter and only examine a slice of the array, where the index represents lsb of the array slice. For example:

for (i=9; i>=0; i=i-1) begin // lsb index of the slice
  if (out[i+:11] >= cnt) begin // 11-bit slice compare
    out[i+:11] = out[i+:11] - cnt; // 11-bit slice subtraction
    avg[i] = 1'b1; // 1-bit assign

This for loop unravels to 10 iterations (9 to 0), each iteration only looks at a 11-bit slice of out and only one bit of avg. You might not be familiar with the +: operator. It is a bit-slice operator introduced in IEEE Std 1364-2001. Left side if the start index (dynamic is allowed) and the right side is the bit with offset (must be a static constant). You can read more about it here.

Since it is a count down, we can can safely assume (proven mathematically) the upper bits of the slice are zeros and we will never have underflow with the guarding if condition. So we now have ten 11-bit subtracters each with 1-bit assigners which is much smaller logic then the original 642 (should be 1024) 20-bit subtracters each with 10-bit adder.

Few other suggestions:

  1. Use ANSI style header (supported since IEEE Std 1364-2001). It is few lines of code and thereby less prone to typos and copy-paste mistakes.
  2. Separate the synchronous logic and combination logic. This does mean declaring more signnals, but generate gives you better control on what is a flop and what is comb logic. It also follows best practices. Your for loop should exits in the comb logic.
  3. Use non-blocking assignments (<=) in your synchronous logic block. This follows best coding practices and prevents flop-to-flop race conditions in simulation.
  4. out can be simplified to a 10-bit register, assuming you did suggestions 2 & 3. This is because we know the upper bits will always be zeros, so we can save 10 flops.

proof of concept: with a simple SystemVerilog testbench here:

module averager( // ANSI style header
  input              clk, rst,
  input [9:0]        n,
  output reg [19:0]  sum,
  output reg [9:0]   cnt,
  output reg [9:0]  out, // was [19:0]
  output reg [9:0]   avg

  reg [19:0] next_sum, next_out;
  reg [9:0]  next_cnt, next_avg;

  integer i;

  always @* begin
    next_sum = sum + n;
    next_out = next_sum;
    next_cnt = cnt + 1'b1;
    next_avg = 10'b0;

    for (i=9; i>=0; i=i-1) begin // lsb index for slice
      if (next_out[i+:11] >= next_cnt) begin // 11-bit slice compare
        next_out[i+:11] = next_out[i+:11] - next_cnt; // 11-bit slice subtract
        next_avg[i] = 1'b1;  // 1-bit assign

  always @(posedge clk) begin
    if (rst == 1'b1) begin
      sum <= 20'b0;
      cnt <= 10'b0;
      out <= 10'b0; // was 20'b0
      avg <= 10'b0;
    else begin
      sum <= next_sum;
      cnt <= next_cnt;
      out <= next_out[9:0]; // only assign lsb
      avg <= next_avg;


If you still experience area issues or performance is not good enough, then you should pipeline your design and/or seeing if there are there is dedicated divider+remainder module defined in your FPGA data sheet that you can instantiate.

  • \$\begingroup\$ ERROR:HDLCompiler:532 - "D:\Users\Averager_new\cma.v" Line 24: Index <10> is out of range [9:0] for signal <next_avg>. I guess it needs some re-indexing. \$\endgroup\$ – vikiboy Apr 11 '15 at 14:31
  • \$\begingroup\$ @vikiboy, I'm surprised I missed that. It should be ten 11-bit sliced subtraction. This is assuming the max sum is 0xFF801 (0x3ff*0x3ff) as design does not have overflow protection. I'll update my answer in a bit \$\endgroup\$ – Greg Apr 11 '15 at 15:51
  • \$\begingroup\$ Okay, the edaplayground link has been updated too with ten 11-bit subtractions. \$\endgroup\$ – Greg Apr 12 '15 at 22:46
  • \$\begingroup\$ well the module went about with a few warnings only. But the test bench had quite a lot of errors. ERROR:HDLCompiler:806 - "" Line 8: Syntax error near "*)". ERROR:HDLCompiler:806 - "" Line 11: Syntax error near "#". ERROR:HDLCompiler:31 - "" Line 11: <n> is already declared. ERROR:HDLCompiler:806 - "" Line 14: Syntax error near "initial". ERROR:HDLCompiler:806 - Line 16: Syntax error near "$monitor". ERROR:HDLCompiler:806 - Line 17: Syntax error near "$dumpfile". \$\endgroup\$ – vikiboy Apr 14 '15 at 17:18
  • \$\begingroup\$ ERROR:HDLCompiler:806 - Line 18: Syntax error near "#". ERROR:HDLCompiler:806 - Line 23: Syntax error near "'". ERROR:HDLCompiler:806 - Line 27: Syntax error near "$finish". ERROR:HDLCompiler:806 - Line 28: Syntax error near "end". ERROR:HDLCompiler:806 - Line 30: Syntax error near "->". ERROR:HDLCompiler:806 - Line 32: Syntax error near "->". ERROR:HDLCompiler:31 - Line 32: <property> is already declared. ERROR:HDLCompiler:598 - Line 3: Module <tb> ignored due to previous errors. \$\endgroup\$ – vikiboy Apr 14 '15 at 17:19

I'm not sure what experience you have in embedded software design (mostly sequential instructions), but a true division takes FOREVER by comparison with just about anything else, even if such an instruction exists. So we have all kinds of tricks to accomplish the high-level goal without actually dividing. Two of them are:

  • Bit-shifting: This is so cheap and easy, it's almost like cheating. It does restrict you to powers of 2, however. Shift-right by N is equivalent to dividing by 2^N.
  • Fixed-point arithmetic: This can be tricky to understand, but the basic idea is that you decide as a developer (the hardware doesn't know or care) to interpret each number as if it had a fractional point between certain bits, as in binary 1011.1001 = decimal 185/16 = 11.5625. (a debugger would read that byte as decimal 185) This allows you to multiply by a reciprocal, which is more expensive than shifting but still cheaper than dividing, and it allows any number that can be stored.
    • Of course, the product may have the fractional point between a different set of bits if you insist on it being mathematically correct, but nothing really says it has to be. At this point, you could conveniently accept the raw product as being scaled up or down automatically, or you could bit-shift it into a different range. Just be careful not to drop any significant bits.
    • And this only works for division by a constant (constant as far as this piece of code is concerned). As far as I know, it takes a true division to generate the reciprocal, and so you're no better off unless you can do that once and reuse the result.

Why not floating-point? Because FP arithmetic in general is about the only thing that's more expensive than integer division. At least the higher-end CPU's have dedicated hardware for it.

I know you're working on a parallel machine and my experience is on a sequential one, but I'm pretty sure they're identical at the level that I just described. My machine reuses the same logic block for each operation in turn, while yours has a dedicated one for each. Just time vs. space.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.