FPGA Intellectual Property
PCI Express*, Networking and Connectivity, Memory Interfaces, DSP IP, and Video IP
6354 Discussions

Specifying timing relative to EMIF pll reference clock

WKapp
Beginner
1,355 Views

Hello all,

I am using an Arria 10 10AX027H4F35E3SG and have included a DDR4 EMIF. I am feeding the FPGA a 133MHz clock as the pll reference. This same clock is used for all the board interface timing.

 

When I compile the FPGA and try to establish timing constraints relative to this input clock, the tools won't let me. It will let me specify timing constraints relative to a clock output from the pll, but that differs from the input clock by several ns.

 

In the auto-generated verilog for the pll instance, it sets the compensation_mode to "emif" and I cannot find any documentation for that mode. It appears from the comments in the verilog that similar to "direct" mode, the pll does not phase the outputs to be in sync with the input clock. I could not find any option in the IP generation that would let me specify the compensation_mode of the IOPLL used in the EMIF.

 

What is the recommended way to constrain timing to a board clock that is used as the reference for the EMIF PLL?

 

Would it be safe to manually edit the pll verilog and change the compensation_mode from "emif" to "normal"?

 

Thanks.

0 Kudos
12 Replies
NurAida_A_Intel
Employee
1,029 Views

Dear WKapp,

 

EMIF mode has been removed or no longer available from the Assignment Editor list in newer Quartus version ( for example v18.1). Hence, may I know what Quartus version are you using? I suggest you to use the latest version like 18.1.

 

Below is just for your information about the EMIF modešŸ˜Š

  • The EMIF mode should only be used by the EMIF IP. In differences perspective between the direct and EMIF mode, both are effectively the same. The different is that the VCO frequency range is slight different. And we suggest user to continue use the direct mode because as mentioned they are going to remove the EMIF mode from the list in newer version as I mentioned above.

 

By default, the clock feedback mode is "direct" compensation mode where PLL does not compensate for any clock networks. To change the mode, you can directly change at the Assignment Editor or your verilog file and re-compile your design.

 

Hope this helps.

 

Thanks

 

Regards,

NAli1

 

 

 

0 Kudos
WKapp
Beginner
1,029 Views

NAli1,

Thanks for the reply. Not sure the source of your information, but my IP was generated with Quartus Prime Pro 18.1 and it has selected "emif" as the compensation_mode for the PLL. I also generated the IP with Quartus Prime Pro 19.1 and it, too, selects "emif" for the compensation_mode.

 

I did try to manually edit the file to select "normal", but this fails in the fitter, unable to place an auto promoted global signal from the PLL.

 

I would like to ask a more general type question:

When designing an FPGA such as the Arria 10 and included in the design is an EMIF such as mine, the PLL outputs the clocks that should be used for all the FPGA logic inside. Now, these PLL output clocks are not phase related to the PLL input reference clock. So, what are the best practices for interfacing the FPGA to other board elements?

 

In my case, I have a board clock that is used by essentially everything on the board. In a previous FPGA (a GX130), its memory controller (an ALTMEMPHY) included a PLL that operated in "normal" mode, so its output clocks were in phase with the input reference clock, and I fed that from the board clock. Now, with the Arria 10, all the interfaces that use the board clock can never be used with the FPGA io since that clock is not in phase with the FPGA fabric.

 

Thanks,

Bill

0 Kudos
NurAida_A_Intel
Employee
1,029 Views

Hi WKapp,

 

I am sorry if I misunderstood your question. Perhaps what I replied previously is too general. Let me try to clarify again and be more specific. šŸ˜Š Letā€™s talk about 2 different perspective as below:

1.     EMIF IP perspective:

2.     General perspective ( Ex: customer own design and etc)

 

1.     EMIF IP:

 

Question: What is the recommended way to constrain timing to a board clock that is used as the reference for the EMIF PLL?

Answer: For Arria 10 EMIF, the clock input pin (emif pll ref clk) is connected to dedicated clock pin thus you do not need additional constrain timing.

 

Question: Would it be safe to manually edit the pll verilog and change the compensation_mode from "emif" to "normal"?

Answer: I assume you are talking about EMIF IP. For EMIF, the compensation mode is already customize for that EMIF IP and it is not changeable. In short, user do not need to worry about what mode they are using because the IP will take care of it.

 

For more information about the PLL clock you may refer to this emif handbook Ć  https://www.intel.com/content/dam/www/programmable/us/en/pdfs/literature/ug/ug-20115.pdf

 

2.     General perspective ( Ex: customer own design and etc)

I am not sure why you still able to see the ā€œemifā€ setting as at my side I am not seeing this listed (see my 2nd screenshot). Even in Arria 10 device handbook chapter 4.2.4. Clock Feedback Modes also didnā€™t support ā€œemifā€ mode. Ć  https://www.intel.com/content/dam/www/programmable/us/en/pdfs/literature/hb/arria-10/a10_handbook.pdf

1.png

Also another way I would recommend you change the PLL compensation mode is by generating the IOPLL from the Qsys IP with the correct compensation mode selected.

2.png

 

Hope this answered the questions.

 

Regards,

NAli1

0 Kudos
WKapp
Beginner
1,029 Views

NAli1,

I am sorry if I am not stating my question properly, I do not think you understand what I am asking. So, I will try to ask more clearly with the help of a picture.

 

Question: What is the recommended way to constrain timing to a board clock that is used as the reference for the EMIF PLL?

Answer: For Arria 10 EMIF, the clock input pin (emif pll ref clk) is connected to dedicated clock pin thus you do not need additional constrain timing.

 

I understand that the memory timing is all constrained, and in my design the memory portion meets timing. My question is not about the memory timing, it is about timing of other FPGA elements. Please look at the following picture (sorry it is rotated):

FPGA_io.jpg

First of all, the clock distribution on the board guarantees that all the clocks are in phase at their destinations. For the purpose of this analysis, please assume the skew between clocks is 0.

 

I have shown 1 input to the FPGA and 1 output from the FPGA, both are connected to external FFs. The board clock that connects to the FFs is the same clock used as the PLL reference clock input. The PLL shown inside the FPGA is the one instantiated as part of the EMIF, and one of its output clocks is used inside the FPGA to clock all the fabric.

 

My problem is that the since the PLL has compenstation_mode = "emif", there is no guaranteed phase relationship between the PLL reference clock input (board clock) and the internal FPGA clock (afi_clk). I want to specify timing constraints for the inputs and outputs relative to the board clock but cannot since there is no known phase relationship.

 

I originally thought I could change the mode of the PLL manually to "normal". If I could, this would make afi_clk in phase with the PLL reference clock, and I could use the PLL reference clock in my timing constraints. However, I cannot set the mode to "normal". It appears that "emif" mode is similar to "direct" mode in that it does not care about the phase.

 

I hope the above is clear.

 

So, please explain, how do I create timing constraints for my example input and output?

 

Thanks,

Bill

0 Kudos
NurAida_A_Intel
Employee
1,029 Views

Hi WKapp,

 

Sorry for the delay in response because we are on long holiday in my region.

 

For the EMIF IP, it should connect to DDR4 module, but we don see DDR4 in the diagram.

 

If I assume the "FF" is referring to DDR4 module, we don't need connect clock to the DDR4 externally. The clock to DDR4 module is supply from FPGA<-->DDR4 interface. 

 

Thanks

 

Regards,

NAli1

 

0 Kudos
WKapp
Beginner
1,029 Views
NAli1, I am afraid you do not understand at all. Let me try again. The diagram shows a box labeled Clock Distribution. This box outputs 3 copies of the exact same clock on the board. All are matched lengths and for this analysis you should assume there is 0 skew between them. One clock goes to an external FF that outputs IN1. One clock goes to external FF whose input is OUT1. The last clock goes to the FPGA and is the pll reference clock input for the EMIF. The pll inside the FPGA is part of the EMIF IP, I did not show the rest of the EMIF nor did I show the DDR4 as they are not important to this discussion. As I am sure you are aware, the EMIF pll outputs the clocks that are to be used by all the user logic inside the FPGA. I have shown just one of these clocks, afi_clk. It is used to clock internal FFs that clock in the signal IN1 and also clock out the signal OUT1. In this design, the frequency of the pll input reference clock is the same as afi_clk. As I am sure you are also aware, the phase of afi_clk is not controlled with respect to the input reference clock. This is the crux of the problem. If the pll were capable of operating in ā€œnormalā€ mode, the output clocks would be in phase with the input reference clock. However, since the pll is in ā€œemifā€ mode, the phase is not controlled. Without any phase relationship, there seems to be no way to write timing constraints for IN1 and OUT1 since the clocks used by the external FFs are not in phase with afi_clk. So, how does one write timing constraints for the FPGA inputs and outputs like IN1 and OUT1? Thanks, Bill
0 Kudos
NurAida_A_Intel
Employee
1,029 Views

Dear WKapp,

 

I really thank you for your detail explanation and now I think I understand your issue better. šŸ˜Š

 

You are trying to constraint timing for other logic (IN1 and OUT1) and not the EMIF as the EMIF donā€™t need the additional timing constraint as per I explained before.

 

And from the diagram you attached, I can see that you are using the afi_clk to drive non-EMIF logic. As far as I know, there are no afi_clk exported from EMIF IP unless you are using ā€œHard PHY onlyā€ mode. Are you using ā€œHard PHY onlyā€ mode?

This afi_clk is actually meant for the custom controller to use and not for other logic like External FF.

 

Anyway, here is some constraint guideline that I hope is helpful for you to get some idea on how to constraint timing  --> https://www.intel.com/content/dam/www/programmable/us/en/pdfs/literature/manual/mnl_timequest_cookbook.pdf

 

I believed the problem you are facing now is more to timing issue and not EMIF. My suggestion to you is to open new thread if you encounter any issue from the documentation and Timing expert will help you.

 

Thanks

 

Regards,

NAli1

0 Kudos
WKapp
Beginner
1,029 Views
NAli1, Yes, I am using Hard PHY only. I know how to write timing constraints. I am not looking for help in how to create the sdc file in detail. What I was hoping to get from this original post was a philosophical answer about how to approach timing in my situation. The design I am working on was a working design but had been in a Stratix II GX part. The memory controller was an ALTMEMPHY. That controller included a PLL but its output clocks were in phase with the PLL reference clock. All the board timing worked out since the board clock was used as the PLL reference clock. The EMIF PLL does not allow me to generate output clocks that are in phase with the PLL reference clock. If I use the PLL reference clock in my timing constraints, there is a huge amount of variation imposed by the fact that the clocks are not in phase, and it becomes difficult if not impossible to meet timing at 133MHz, something that was childā€™s play for the Stratix II GX. I was hoping you knew something I did not, and there was a trick to help me with timing in this situation. I canā€™t be the first one that has implemented an EMIF and needs a board clock, too! I am not in a position to put FIFOs in all the interfaces with different clocks as latency for these interfaces is important. Am I missing something? Bill
0 Kudos
NurAida_A_Intel
Employee
1,029 Views

Hi WKapp,

 I consult timing expert guy, and he also aware about the difficultly in meeting timing based on your situation as the clock is not in the same phase.

His recommendation for you is to connect the afi_clk out externally to your External FF instead of using clock distribution so that all the clock will have same source be in same phase.

I attached diagram for your reference.

Timing.PNG

Hope this helps.

 

Thanks

 

Regards,

NAli1

 

0 Kudos
WKapp
Beginner
1,029 Views
NAli1, Thank you for the reply. I had pretty much figured out there was no solution for this situation. I am surprised the designers of the IP did not consider this situation and accommodate it. Oh well, canā€™t change it nowā˜¹. Unfortunately, I cannot use afi_clk externally as this FPGA is not the source of the board clock. There are many other FPGAs and external logic that all must be clocked in phase. Thanks, Bill
0 Kudos
NurAida_A_Intel
Employee
1,029 Views

Dear WKapp,

 

Understand the pain that you are facing. I sincerely apologize for the inconvenience caused.

 

Regards,

NAli1

0 Kudos
WKapp
Beginner
1,029 Views
No need to apologize. Thanks for helping. Bill
0 Kudos
Reply