Possible bug in Vertex-4 Rocket-IO?

I need to use Rocket-IO in Vertex-4 to receive data at the rate of
6.22Gbps. My reference clock input to the Rocket-IO is 155.5MHz. Data
rate is 40X of that clock. I need to have a parallel output in
40-bit-wide format and output at 155.5MHz rate. I do not need any other
process inside the Rocket-IO. I just need to do serial-to-parellel
conversion. I have followed the guide of Rocket-IO and also got some
help from Xilinx engineer to get a start. Now I have simulated with VCS
to see the Rocket-io output data as I intended to.

However if I run the simulation longer I see the parallel output become
incorrect at every 40th parellel output word. The expected value is not
there. Instead the value in the previous clock cycle is repeated. This
goes on at every 40th word. All other output words are correct.

I am wondering if any other V4 Rocket-IO user faces the same problem.
Could it be the simulation model's problem, Xilinx Rocket-IO's bug, or
something I did not set up correct? Unfortunately the same Xilinx
engineer keeps quite since I sent my question to him. Maybe too busy?

Appreciated for any help. If needed I can send all the files including
instantiation to GT11 and testbench. These are verilog files.

What clock are you using to clock the data out?  It should not be your
reference clock because this will not be synchronous to the data and
you will experience a glitch periodically.  You could use the
recovered clock.  This is the clock recovered from the serial data
stream so it will be synchronous with the data.  

On 7 Jul 2005 23:44:21 -0700, shuo.huang@fibre.com wrote:

>I need to use Rocket-IO in Vertex-4 to receive data at the rate of
>6.22Gbps. My reference clock input to the Rocket-IO is 155.5MHz. Data
>rate is 40X of that clock. I need to have a parallel output in
>40-bit-wide format and output at 155.5MHz rate. I do not need any other
>process inside the Rocket-IO. I just need to do serial-to-parellel
>conversion. I have followed the guide of Rocket-IO and also got some
>help from Xilinx engineer to get a start. Now I have simulated with VCS
>to see the Rocket-io output data as I intended to.
>
>However if I run the simulation longer I see the parallel output become
>incorrect at every 40th parellel output word. The expected value is not
>there. Instead the value in the previous clock cycle is repeated. This
>goes on at every 40th word. All other output words are correct.
>
>I am wondering if any other V4 Rocket-IO user faces the same problem.
>Could it be the simulation model's problem, Xilinx Rocket-IO's bug, or
>something I did not set up correct? Unfortunately the same Xilinx
>engineer keeps quite since I sent my question to him. Maybe too busy?
>
>Appreciated for any help. If needed I can send all the files including
>instantiation to GT11 and testbench. These are verilog files.

I used RXRECCLK to clock the output.

Marko wrote:
> What clock are you using to clock the data out?  It should not be your
> reference clock because this will not be synchronous to the data and
> you will experience a glitch periodically.  You could use the
> recovered clock.  This is the clock recovered from the serial data
> stream so it will be synchronous with the data.  
>

So, you are getting more clocks output than you are inputting?  This
must be the case if you are repeating codes.  I guess the other
possibility is that you are inadvertently sending the same code twice
and your receiver is working properly. 


On 8 Jul 2005 08:29:21 -0700, shuo.huang@fibre.com wrote:

>I used RXRECCLK to clock the output.
>
>Marko wrote:
>> What clock are you using to clock the data out?  It should not be your
>> reference clock because this will not be synchronous to the data and
>> you will experience a glitch periodically.  You could use the
>> recovered clock.  This is the clock recovered from the serial data
>> stream so it will be synchronous with the data.  
>>

The number of parallel output words is correct. I do not get more
clocks at output. Every 40th word's value is replaced by the 39th. I
have checked the serial data. It is correct. It is the Rocket-IO that
works incorrect.

Marko wrote:
> So, you are getting more clocks output than you are inputting?  This
> must be the case if you are repeating codes.  I guess the other
> possibility is that you are inadvertently sending the same code twice
> and your receiver is working properly.

So, you get word 1-38, 39, 39, 41, 42 ... Is that correct?
You repeat word 39 and skip word 40?

Is the problem data dependent?  If you change the data, is the effect
the same?

Do you have a comma alignment character embedded?

Is this a single channel application or are you bonding multiple
channels together?



On 9 Jul 2005 09:05:15 -0700, shuo.huang@fibre.com wrote:

>The number of parallel output words is correct. I do not get more
>clocks at output. Every 40th word's value is replaced by the 39th. I
>have checked the serial data. It is correct. It is the Rocket-IO that
>works incorrect.
>
>Marko wrote:
>> So, you are getting more clocks output than you are inputting?  This
>> must be the case if you are repeating codes.  I guess the other
>> possibility is that you are inadvertently sending the same code twice
>> and your receiver is working properly.

The problem is not data dependent. I did not enable comma alignment.
This is a single channel application.

Did you have succesfull experience with similar settings?

Forgot to confirm your first paragraph. Yes. That is the pattern I see
in the simulation.

Thanks for your help.

I had good luck with Rocket IO in Virtex2Pro.  I used comma alignment,
bonded two channels together and use the recovered clock to clock
everything in the receiver (except the refclk).  My bit rate was only
about 1Gbps.  In my case, everything worked in simulation and in
hardware.

If you don't use comma alignment, how to you align the serial stream
to the word boundary?  Are you using a standard protocol or custom?

You say this happens in simulation.  I presume you don't have
hardware?  I presume you purchased the smart model extension for
ModelSim?  If so, can you get support from ModelTech and/or Xilinx on
this issue?  Have you followed all the rules for proper simulation.
One rule I remember was that simulation must run at pico-second speed.

Sorry I can't be of more help.


On 9 Jul 2005 14:35:40 -0700, shuo.huang@fibre.com wrote:

>The problem is not data dependent. I did not enable comma alignment.
>This is a single channel application.
>
>Did you have succesfull experience with similar settings?

Thanks Marko for sharing your experience.

I do not use comma alignment inside the Rocket-IO. I have circuit after
it to do the alignment. I am planning to use Virtex-4 device (if it
works). I do not think there is any hardware available yet.

I used VCS (from Synopsys) to simulate. It takes smartmodel as well. I
am trying to get help from Xilinx for a number of weeks. But I have not
got any helpful solution yet.

I noticed that in Roket-IO model the time scale (time unit definition
in Verilog) is defined in ps. I tried to modify my test bench from ns
to ps. Nothing was changed. In Verilog it is allowed to have different
time scales in different modules. I think the simulation with
smartmodel settings should be correct, because otherwise it would not
work at all. It would not be only at every 40th output.

Thanks for your time. I will post the final result when I have it.