How to buffer data into MEMTILE and read it multiple Times?

[ngdymx]

Hi team,

Is there any way that I can buffer data into MEMTILE and read it multiple times?
Thanks a lot!

[AndraBisca]

To do that, you can set the repeat_count input of the AIE_DMAStartOp to a value >= 1, depending on how many more times you will read the data. This will enable the DMA associated with your data movement to push a new copy of the data on the stream.

I am currently working on adding a feature to the AIE_ObjectFifoCreateOp to be able to describe this data repeat, however, the PR isn't ready just yet. I am aware of some examples without object FIFOs that currently use the repeat feature here.

Hope this helps!

[ngdymx]

Hi,

Thank you so much! I will try it.

[ngdymx]

Hi,

I hope you are doing well.

I wanted to inquire about the current status of adding repeat count to objectFifo. I understand you might be busy, and I do not mean to rush you. However, I wanted to let you know that I am still awaiting it.

Thank you for your time and understanding.

[AndraBisca]

Hello,

Thank you for your patience. Unfortunately, while we have made some progress with testing this feature, it has been a little slow due to other ongoing work on different features and fixes. However, I would be curious to know whether you've run any tests related to buffering in the memtile without access to the feature for now? Any additional input is welcome to help speed up its implementation.

Thank you for your understanding and also for checking in on this discussion so that it stays up to date!

[ngdymx]

Hello,

I tried to run the test_repeat_count.py, but it gave me the cannot import name 'linalg' from 'aie.extras.dialects.ext' error. Later, I tried to mimic the code but it failed. ☹️

Then I tried to buffer the data into DDR and read it multiple time, showing in the below code. It worked only if I put "width" in the fourth dimension. If I put the "width" in the second or third dimension, it will fail.

npu_dma_memcpy_nd(
    metadata="inOF_wts_0_L3L2",
    bd_id=1,
    mem=W,
    offsets = [0, 0, 0, 0],
    sizes=[width, 1, 1, weightsInInt32s],
    strides=[0, 0, 0],
 )

So far, I haven't successfully buffered the data in the MEMTILE.