Page 1 of 2
Questions on ethernet (XE devices) and handing data reads
Posted: Wed Jul 20, 2016 11:57 am
by DemoniacMilk
Hello!
For our current project we are using an XE216 on an eXplorerKIT.
We would like to implement USB, Ethernet, SPI, SSC and Audio sampling/compression.
Because the xcore doesnt have DMA, each communication interface needs a core to sample input data?
From what i understood, USB requires 2 cores (XUD, EP0) plus two cores to remain idle (correct?), SPI needs an additional core, SSC will probably be implemented on 3 cores (im a newbie so i dunno yet. Looking to go for an RX,TX,Manager setup I think). Audio sampling and processing (low quality mono audio) should be possible on a single core. Gigabit ethernet consumes 7+1 cores, for a total of 17 cores, plus application threads.
--> not enough ressources available on XE216.
For the target application, gigabit ethernet is not needed. 100 Mbit ethernet is sufficient. Is it possible to use a mii instead of an rgmii interface on xe216 devices/the eXplorerKIT (Id guess no, cause its two different interfaces)? Do you think theres a way of freeing ressources in this setup?
Re: Questions on ethernet (XE devices) and handing data read
Posted: Wed Jul 20, 2016 1:48 pm
by peter
Hi DemoniacMilk,
There are a few things that you could do, but they are clearly beyond what XMOS tests internally and therefore I can't guarantee they will work.
The first is that you could run the XUD as high priority to allow use of all of the 7 other cores on that tile. This
Application Note describes how to do that.
The second is that you can restrict the RGMII interface to only run at 100Mbit which would limit it to
a maximum of 5 cores. However, this would require modification of the sources in lib_ethernet.
You need to ensure that you only request 100MBit operation:
Code: Select all
smi_configure(smi, phy_address, LINK_100_MBPS_FULL_DUPLEX, SMI_ENABLE_AUTONEG);
And then you'll need to replace all of the contents of the 'else' at
line 215 of rgmii_ethernet_mac.xc with something like:
Code: Select all
assert(0); // RGMII no longer supported.
If you just do that you'll get 1 extra core. In order to get down to 5 cores you need to change:
Code: Select all
{
rgmii_buffer_manager(c_rx_to_manager[0], c_speed_change[3],
*p_used_buffers_rx_lp, *p_used_buffers_rx_hp, *p_free_buffers_rx, 0);
}
{
// Just wait for a change from 100Mb mode and empty those channels
c_speed_change[2] :> unsigned tmp;
c_speed_change[4] :> unsigned tmp;
}
to:
Code: Select all
{
rgmii_buffer_manager(c_rx_to_manager[0], c_speed_change[3],
*p_used_buffers_rx_lp, *p_used_buffers_rx_hp, *p_free_buffers_rx, 0);
// Just wait for a change from 100Mb mode and empty those channels
c_speed_change[2] :> unsigned tmp;
c_speed_change[4] :> unsigned tmp;
}
Hope that helps,
Peter
Re: Questions on ethernet (XE devices) and handing data read
Posted: Wed Jul 20, 2016 1:50 pm
by peter
Thinking about it, you might need to use
instead of
in order to prevent it getting into gigabit mode.
Re: Questions on ethernet (XE devices) and handing data read
Posted: Thu Jul 21, 2016 9:23 am
by DemoniacMilk
Hello Peter,
thank you for your awesome answer!
Ethernet is now running off 5 cores instead 7 and processes ICMP requests just fine! (although I dont understand how changing
Code: Select all
{
rgmii_buffer_manager(c_rx_to_manager[0], c_speed_change[3],
*p_used_buffers_rx_lp, *p_used_buffers_rx_hp, *p_free_buffers_rx, 0);
}
{
// Just wait for a change from 100Mb mode and empty those channels
c_speed_change[2] :> unsigned tmp;
c_speed_change[4] :> unsigned tmp;
}
affects the program).
I enabled high core priority in xud_(), is that a reasonable place to enable it in?
Re: Questions on ethernet (XE devices) and handing data read
Posted: Thu Jul 21, 2016 9:39 am
by peter
The change in the ethernet is simply merging two parallel tasks into one. There was no need for them to be separate before, but since the RGMII was taking 7 cores there was also no need to change anything to reduce the core count for 10/100 ethernet case.
In terms of the USB, the place to do it is as the document says for the XUD_Manager task:
Next modify the file module_usb_audio\main.xc, as shown below :
Code: Select all
/* USB Interface Core */
{
set_core_high_priority_on();
#if (AUDIO_CLASS==2)
XUD_Manager(c_xud_out, ENDPOINT_COUNT_OUT, c_xud_in, ENDPOINT_COUNT_IN,
c_sof, epTypeTableOut, epTypeTableIn, p_usb_rst,
clk, 1, XUD_SPEED_HS, XUD_PWR_CFG);
#else
XUD_Manager(c_xud_out, ENDPOINT_COUNT_OUT, c_xud_in, ENDPOINT_COUNT_IN,
c_sof, epTypeTableOut, epTypeTableIn, p_usb_rst,
clk, 1, XUD_SPEED_FS, XUD_PWR_CFG);
#endif
}
This configures the core running the XUD to high priority to ensure that it has enough MIPS to execute
the USB interface. Setting this will allocate 100MIPS to the USB task, leaving 400MIPS to be shared equally
among the remaining tasks. It allows all remaining logical cores to be utilized. 400/7 = 57MIPS is
sufficient for all other functions within the reference design when running at up to 384KHz, with stereo
out.
Re: Questions on ethernet (XE devices) and handing data read
Posted: Tue Nov 08, 2016 7:30 pm
by mozcelikors
Hello, I also wanted to reduce ethernet core usage, so I went and did what is written. However, I get the following error in console:
Code: Select all
xrun: Program received signal ET_ECALL, Application exception.
[Switching to tile[1] core[0]]
rgmii_ethernet_mac (i_rx_lp=@0x7fd94, i_tx_lp=<value optimized out>, c_rgmii_cfg=<value optimized out>, rgmii_ports=<value optimized out>) at C:/Users/MyUser/Documents/XTimeComposerStudio_Workspace3/lib_ethernet/src/rgmii_ethernet_mac.xc:227
227 assert(0); //RGMII No longer supported in this version
Re: Questions on ethernet (XE devices) and handing data read
Posted: Tue Nov 08, 2016 8:16 pm
by mozcelikors
Update: Changing
Code: Select all
rgmii_inband_status_t current_mode = INITIAL_MODE;
to
Code: Select all
rgmii_inband_status_t current_mode = INBAND_STATUS_100M_FULLDUPLEX_UP;
worked.
Thanks for the helpful thread.
Re: Questions on ethernet (XE devices) and handing data reads
Posted: Wed Apr 19, 2017 5:16 pm
by akp
A note from my implementation.
1.
Code: Select all
rgmii_inband_status_t current_mode = INBAND_STATUS_100M_FULLDUPLEX_DOWN;
Seems to make more sense than an initial state of INBAND_STATUS_100M_FULLDUPLEX_UP, and it also works.
2.
I would recommend using
Code: Select all
smi_configure(smi, phy_address, LINK_100_MBPS_FULL_DUPLEX, SMI_ENABLE_AUTONEG);
This appears to disable autonegotiation to gigabit but retains autonegotiation to 10/100. When I used SMI_DISABLE_AUTONEG I had to force the other end of the link to 100Mbps in order to get the connection to work (i.e. I couldn't leave the other end of the link in automatic mode, so SMI_DISABLE_AUTONEG is a little inconvenient for the end user).
Caveat is that so far I have tested the link with the following setup only: directly connected to a Mac that has a Gigabit, AVB capable NIC. It negotiates to 100 Mbps and my core usage in AVB mode is down from 8 to 6 on tile 1 of the xCORE-200 Multichannel Audio Platform 2v0
Like the OP I only need 100Mbps Ethernet, I don't need gigabit. So this leaves free cores for adding in some other features I am thinking about.
Re: Questions on ethernet (XE devices) and handing data reads
Posted: Fri Aug 11, 2017 4:17 pm
by ast8346
Hello,
I have a problem modifying the contents of the 'else' at line 215 of rgmii_ethernet_mac.xc. Removing content of this declaration generates the following error:
Error: Undefined reference to '_Srgmii_buffer_manager_0'
The code is simply that:
else{assert(0);}
No problem with modifying other bits, but that one is the most important...
Re: Questions on ethernet (XE devices) and handing data reads
Posted: Tue Nov 26, 2019 11:06 pm
by akp
As an update, it should be possible to combine rgmii_ethernet_rx_server and rgmii_ethernet_tx_server into a single task. That would reduce your core usage by one. I don't know if it will work in Gbps mode, so it might be possible in 100 Mbps mode only.