HLS core seems to ignore configuration s_axilite arrays

Hi Folks,
I’ve written (and verified with csim and cosim) and HLS core that performs a multichannel direct digital downconversion operation. The core accepts an axi stream of 8 complex ap_fixed<16,-9> numbers (256 bits) and outputs a stream of 8 complex ap_fixed <-16,7> numbers (also 256 bits). The streams are 256 TDM, so I’ve effectively got 2048 DDCs.

Configuration is via two arrays (ap_fixed<16,1> toneinc[256][8] and ap_ufixed<16,1> phase0[256][8]) used to set the phase accumulator increment and offset. Arrays use an s_axilite interface.

Each transaction the core loads the appropriate 8 increments, phase offsets, and stored phases, computes the current phases, queries sin cos LUTs to get complex numbers, does a complex multiply on the input samples. Finally it stores the updated accumulated phases.

The trouble I’m running into is that I can’t seem to reliably get the core to work from python. It seems that my driver writes and reads the toneinc and phase0 arrays properly but they do not seem to reliably impact the data that moves through the core.

For example I’d expect to be able to send 20 packets of data (20x2048) with real only values, 0 tone inc, 0 phase0, change to 0.5 phase0, send 20 packets, change to .0 phase0, and send 20 packets again. If I’m sending the same data each time I’d expect to see the output go from out=in to out.real=in.imag & out.imag=in.real to out=-in. This doesn’t happen.

I’ve attached a tcl block design from 2019.2, the exported HLS IP, the HLS project source, and an ipython notebook that highlights the issue.

To run the code I added the nice FpBinary library to my pynq image via pip install git+ (592.5 KB)

This is on the ZCU111 (though I doubt that is particularly relevant).

Edit: This new notebook makes it a bit clearer: test_ddc.ipynb (156.0 KB) and eases use with the SystemILA

I checked the notebook; the write_group() call is suspicious. Maybe you want to verify the data you write into AXI lite interface are the ones you want. Also, have you tried to read the AXI lite registers after you write them? Try to check that as well to see if data is consistent.

I hadn’t hooked the axi connection to the ILA and looked at it but I have now. The memories appear to be written with the correct values at the correct addressed and subsequent reads yield the correct data.

I’m increasingly convinced (against all odds) that there is some sort of HLS issue I’m running into. I’ve packaged things up and posted about it here on the Xilinx HLS forums.

The short of it is that if I send the test vector into the core in my implemented design from pynq and ensure it arrives correctly with the ILA the output (as verified with the ILA) does not match the output from C or RTL simulation in HLS. I am at a complete loss.