Forum Discussion

lcy2000's avatar
lcy2000
Icon for New Contributor rankNew Contributor
8 months ago
Solved

readyLatency of Arria 10 PCIe Avalon-ST IP

Hello,

I'm writing a design with Arria 10 AVST PCIe IP (Quartus 22.4; PCIe IP version 20.1.1). The SoC is written in verilog rather than Block Design to promote portability. So we need to handle readyLatency ourself. Also, we use custom 250Mhz clock `soc_clk` to drive `pld_clk` rather than default `coreclkout_hip` to avoid CDC.

And I have encounterd a weird problem regarding ready & valid handshake of AVST interface on the RX path.

Basically , using SignalTap, I saw AVST's valid asserted by the IP after two cycles when ready is asserted, recovering from a backpressure on the RX path. According to AvalonST Spec, valid can only be asserted on ready cycles (n + <readyLatency>). So I believe this implies a readyLatency = 2 on RX path.

But after I checked the pcie_ep.xml, which I believe is a document regarding current IP settings. And I saw readyLatency on both RX and TX path is 3.

So I checked PCIe IP's User Guide (version 18.0, latest on website), and it says that RX path has a readyLatency of 3, but for TX path, it has a readyLatency of 2.

Based on my obverstaion, I'm very confused now. What is the actual ready latency for these path? Is there a reliable way that I can confirm for each IP generation?

Pcie
  • VenT_Altera's avatar
    VenT_Altera
    8 months ago

    Hi Chenyang,


    Thank you for sharing your detailed observations regarding the signal behavior of rx_st_ready and rx_st_valid.


    You are right about the current IP settings in which both the RX and TX path has a readyLatency of 3. I observed the same from our Design Example generated from Quartus.

    According to the User Guide, when it mentions a ready latency of 3 cycles, it does not mean the readyLatency must be exactly 3 cycles. Instead, it means the readyLatency can be up to 3 cycles. The user's Application Layer should be designed to handle the readyLatency up to the specified number, which is 3 in this IP.


    Another way to check your readyLatency is via Quartus Platform Designer.

    1. Open the .qsys file.

    2. Go to View > Component Instantiation.

    3. Locate the Arria 10/Cyclone 10 Hard FPGA PCIe IP.

    4. Under Component Instantiation tab > Signals & Interfaces sub-tab > click rx_st signals.

    5. You should be able to see the Ready latency value configured.


    Arria® 10 and Cyclone® 10 GX Avalon® Streaming Interface for PCI Express* User Guide: https://www.intel.com/content/www/us/en/docs/programmable/683647/18-0/datasheet.html


    I apologize again for my late response. I hope this has clarified your questions regarding the readyLatency of this IP.


    Please let me know if anything remains unclear.


    Thanks.

    Best Regards,

    Ven


6 Replies

  • lcy2000's avatar
    lcy2000
    Icon for New Contributor rankNew Contributor
    8 months ago

    Bump this thread a bit since 10 days have gone.


    Currently after checking example design exported from Quartus, we tried to implement our Application Layer that are capable receiving more than 3 beat of data after ready=0 on RX side (readyAllowance>3), and never assert valid=1 when ready=0 on TX side (readyLatency=0). For now I believe it works and I never saw garbage TLP from HIP, which I believe is an internal buffer overwrite due to incorrect readyLatency.

    The documentation is still unclear. If possible, please check and clarify the readyLatency of this IP. Thank you very much.

  • VenT_Altera's avatar
    VenT_Altera
    Icon for Frequent Contributor rankFrequent Contributor
    8 months ago

    Hi Chenyang,


    My apologies for missing your forum post.

    Please allow me some time to investigate, and I will get back to you in 1-2 days.


    Thanks.

    Best Regards,

    Ven


  • VenT_Altera's avatar
    VenT_Altera
    Icon for Frequent Contributor rankFrequent Contributor
    8 months ago

    Hi Chenyang,


    Thank you for sharing your detailed observations regarding the signal behavior of rx_st_ready and rx_st_valid.


    You are right about the current IP settings in which both the RX and TX path has a readyLatency of 3. I observed the same from our Design Example generated from Quartus.

    According to the User Guide, when it mentions a ready latency of 3 cycles, it does not mean the readyLatency must be exactly 3 cycles. Instead, it means the readyLatency can be up to 3 cycles. The user's Application Layer should be designed to handle the readyLatency up to the specified number, which is 3 in this IP.


    Another way to check your readyLatency is via Quartus Platform Designer.

    1. Open the .qsys file.

    2. Go to View > Component Instantiation.

    3. Locate the Arria 10/Cyclone 10 Hard FPGA PCIe IP.

    4. Under Component Instantiation tab > Signals & Interfaces sub-tab > click rx_st signals.

    5. You should be able to see the Ready latency value configured.


    Arria® 10 and Cyclone® 10 GX Avalon® Streaming Interface for PCI Express* User Guide: https://www.intel.com/content/www/us/en/docs/programmable/683647/18-0/datasheet.html


    I apologize again for my late response. I hope this has clarified your questions regarding the readyLatency of this IP.


    Please let me know if anything remains unclear.


    Thanks.

    Best Regards,

    Ven


  • VenT_Altera's avatar
    VenT_Altera
    Icon for Frequent Contributor rankFrequent Contributor
    7 months ago

    Hi Chenyang,


    May I know if you have any further questions about my previous message?


    Thanks.

    Best Regards,

    Ven


    • lcy2000's avatar
      lcy2000
      Icon for New Contributor rankNew Contributor
      7 months ago

      Hi ventt,


      Thank you for handling this case. I now understand the handshake paradigm of avalon stream.

  • VenT_Altera's avatar
    VenT_Altera
    Icon for Frequent Contributor rankFrequent Contributor
    7 months ago

    Hi Chenyang,


    Thank you for your confirmation.

    I am glad that your questions have been addressed.


    With that, I will transition this thread to community support. If you have a new question, feel free to open a new thread to get support from Altera experts.


    Thanks.

    Best Regards,

    Ven