Peripheral to Memory DMA Transaction using no-OS Driver for axi_dmac HDL IP with ADAQ23878

Hello,

I moved your thread to the FPGA Reference Designs group for better assistance regarding the HDL side.

As we are looking into this issue, could you please provide us more details:

• Is there any particular reason why you chose to use Vivado 2020.2? We offer support only for the latest releases, with their correspondent Vivado versions, that can be found here: Releases · analogdevicesinc/hdl.
• Did you modify the adaq2387 HDL design (except for porting it to another carrier)?

Regards,

Stanca

Hi Stanca,

I was having some UI bugs with the newer Vivado versions, but if you believe the issue is related I can try updating it.

The Vivado project is a simplified version of CN0577. The AXI LTC2387 and AXI DMAC source have not been modified in any way, just cloned the repo and built them.

Kind regards,

Vitor.

Regarding the DMA issue, someone from our team is looking into it and will come back with an answer in the upcoming days.

In the meantime, I have an auxiliary question:

By "simplified version" you mean that you deleted the logic that you found unnecessary from cn0577 and followed the build_hdl guide, or you manually added the IPs in your design either from the Vivado GUI or from a custom tcl script?

I have built the cn0577 project and the IP blocks using the build_hdl guide and used it as a reference. The adaq23878_eval project whose diagram I showed was made by manually adding the IP blocks using the GUI. I also replicated the IP customization and constraints as much as I could.

So I checked out the 2023_r2_p1 HDL release, regenerated all the HDL using Vivado 2023.2, exported and regenerated adaq23878_project for Vivado 2023.2 with write_project_tcl.

I noticed the axi_dmac from 2023_r2_p1 does not have Autorun. I disabled cyclic mode and used the no-os-driver project I shared before with the 2023_R2 branch of the no-OS repo. 

Still got the same result. I am able to see the ADC test pattern 0x330fc in the ILA but no transfer is performed with axi_dmac_transfer_start / axi_dmac_transfer_wait_completion. Destination data is still zero.

Found the problem. We need to invalidate the data cache after the DMA transfer or disable it a the start. 

#include "xil_cache.h"

// Either disable the data cache at startup or invalidate after each DMA transaction
//Xil_DCacheDisable();

/* ... other code ...*/

dma_transfer_and_wait(&transfer1);
Xil_DCacheInvalidate();

This seems to work with both the newer axi_dmac in Autorun/Cyclic mode and the older core (2023_r2_p1 release) in single transfer mode.

I saw there is a cache-coherency option and tried to enable it in the block but it didn't work. Tried the default values and experimented with others but so far only the software solution worked.

Hello bleudoutremer ,

Sorry for the long wait. We discussed internally and the feedback is that we currently support Cache-Coherency only on Linux. We are planning to add support on No-OS, too, but we don't have a specific timeline for that.

Was this the only issue that you encountered?

Hi Stanca,

No problem. That sounds good. Yes, this has been the only issue so far.

I have another question regarding IRQ generation. The docs say that the TRANSFER_COMPLETED interrupt is not generated for cyclic mode.

Is it possible somehow to enable the transfer done interrupts for cyclic transfers, or use another interrupt to monitor the end of transfer?

I was thinking of enabling autorun and setting the transfer size to something like 1K ADC samples so that would give a lower interrupt rate even on cyclic mode.

Hi,

Cyclic mode does not generate interrupts, so you'd have to use another approach to get the end-of-transfer interrupts.

You can either submit the same transfer in a loop, therefore having access to each transfer interrupt individually, or you can use scatter-gather transfers in which you can set the end-of-transfer interrupt flag of each descriptor. Please note that we currently offer support for scatter-gather transfers on linux only.

Regards,

Ionut

Hi Ionut,

OK then, I'll try to use scatter-gather mode to do it.

Will mark this as answered.

Kind regards,

Vitor.