CNN Vivado Build Error

[rcetin]

Hi EJ,

I am trying to build CNN example. I have used this ready-to-use model. ../../hls4ml/keras-to-hls/example-keras-model-files/KERAS_conv1d.json

When i run this command: vivado_hls -f build_prj.tcl
I got this error:

***** C SIMULATION *****
INFO: [SIM 211-2] *************** CSIM start ***************
INFO: [SIM 211-4] CSIM will launch GCC as the compiler.

   Compiling ../../../../test_cnn_test.cpp in debug mode
   Compiling ../../../../firmware/test_cnn.cpp in debug mode
   Generating csim.exe
@E Simulation failed: SIGSEGV.
ERROR: [SIM 211-100] CSim failed with errors.
INFO: [SIM 211-3] *************** CSIM finish ***************
4
    while executing
"source [lindex $::argv 1] "
    ("uplevel" body line 1)
    invoked from within
"uplevel \#0 { source [lindex $::argv 1] } "

INFO: [Common 17-206] Exiting vivado_hls at Wed Sep 18 00:21:57 2019...

Then i tried to remove CSIM from project in build_prj.tcl by commenting these lines:

if {$opt(csim)} {
  puts "***** C SIMULATION *****"
  csim_design
}

After trying build again, i got this error:

***** C/RTL SYNTHESIS *****
INFO: [HLS 200-10] Analyzing design file 'firmware/test_cnn.cpp' ...
ERROR: [HLS 200-70] pragma 'STREAM variable=logits1 complete depth=1' has unknown option 'complete'
ERROR: [HLS 200-70] '#pragma HLS STREAM variable=logits1 complete depth=1' is not a valid pragma.
ERROR: [HLS 200-70] pragma 'STREAM variable=logits2 complete depth=1' has unknown option 'complete'
ERROR: [HLS 200-70] '#pragma HLS STREAM variable=logits2 complete depth=1' is not a valid pragma.
ERROR: [HLS 200-70] pragma 'STREAM variable=logits3 complete depth=1' has unknown option 'complete'
ERROR: [HLS 200-70] '#pragma HLS STREAM variable=logits3 complete depth=1' is not a valid pragma.
Pragma processor failed: 
    while executing
"source [lindex $::argv 1] "
    ("uplevel" body line 1)
    invoked from within
"uplevel \#0 { source [lindex $::argv 1] } "

INFO: [Common 17-206] Exiting vivado_hls at Wed Sep 18 00:25:31 2019...

I can build keras_1layer successfully. What can be a problem about CNN?

Best regards.

[rcetin]

Hi,

I removed the complete keyword from firmware/test_cnn.cpp file. Then, it passed this error but it gave another one:

ERROR: [XFORM 203-504] Stop unrolling loop 'ConvOut' (/home/rcetin/rfnoc/rfnoc-hls-neuralnet/rfnoc/hls4ml/nnet_utils/nnet_conv.h:112) in function 'nnet::conv_1d<ap_fixed<18, 8, (ap_q_mode)5, (ap_o_mode)3, 0>, ap_fixed<18, 8, (ap_q_mode)5, (ap_o_mode)3, 0>, config3>' because it may cause large runtime and excessive memory usage due to increase in code size. Please avoid unrolling the loop or form sub-functions for code in the loop body.
ERROR: [HLS 200-70] Pre-synthesis failed.
command 'ap_source' returned error code
    while executing
"source [lindex $::argv 1] "
    ("uplevel" body line 1)
    invoked from within
"uplevel \#0 { source [lindex $::argv 1] } "

INFO: [Common 17-206] Exiting vivado_hls at Wed Sep 18 15:21:20 2019...

Regards.

[ejk43]

Hi @rcetin

A couple things with the CNN layer:

1. It's a different layer type, so the core HLS code will look very different than the dense layers.
2. You'll want to check out the core hls4ml repo to see if it works in the most recent tree (https://github.com/hls-fpga-machine-learning/hls4ml). There's a TON of updates in the past year and many features are fixed/improved by now.
3. Unfortunately, I have not been keeping the hls4ml submodule in this repo updated. Please feel free to update the git submodule (git submodule update --remote), see if more things work, and submit a PR -- I'm afraid I cannot commit on a timeframe to debug personally, so your best luck here might be to plug in to the hls4ml project for the HLS generation, and then use this repo to help integrate into RFNoC when you're ready for that.

[rcetin]

Hi @ejk43 ,

I have opened an issue in hls4ml repo and got good feedback. In the last news, there are two problems.

1. io_serial is buggy.
2. hls4ml does not support big models yet.

You can find the issue in here

So, if one wants to implement a bigger model, he should wait for hls4ml support or implement himself. However, for a smaller CNN model, I guess I could build the model in Vivado. But it is failed in Xsim. But I think it is not a big problem for now. (Its memory issues again.) You can also find the build log in the issue.

Best regards.