During the writing of an HLS project for a model with a skip connection, only one of the two copies of the stream are used in the first "branch". The original stream is propigated to the second "branch" instead of the second copy. This results in reading from a stream that is already empty.
Commit hash: 107589fb1dac9aa898baea328acc6d519596e19d
TF Model architecture
In addition tot this, the concatenate1d() function of nnet_merge_stream.h has a error in both loops where the out_data assignments don't have an additional offset to account for the progress of the outer loop. The fixed version is given below.
Details
The appropriate output stream is cloned with nnet::clone_stream<>() (layer13_out stream becomes layer71_cpy1 & layer71_cpy2). layer71_cpy1 is then correctly passed to the first "branch" but layer13_out (which has already been read from for the cloning process) is incorrectly passed to the second branch instead of layer71_cpy2. These warnings are produced as a result along with incorrect predictions.
WARNING: Hls::stream 'layer13_out' is read while empty, which may result in RTL simulation hanging.
WARNING: Hls::stream 'layer71_cpy2' contains leftover data, which may result in RTL simulation hanging.
Steps to Reproduce
Download & unzip files.zip
conda env create -f environment.yml
conda activate bepfm_fpga_temp
Install main branch of hls4ml
a. git clone https://github.com/fastmachinelearning/hls4ml.git
b. cd hls4ml;pip install .;cd ..
python3 compile_prj.py
Navigate to sho_fitter_stream_rf8_TEMP/firmware/myproject.cpp
Expected behavior
layer13_out is cloned on line 173 of myproject.cpp to layer71_cpy1 and layer71_cpy2.
layer71_cpy1 should be passed into the pooling1d_cl() on line 177 which works currently.
layer71_cpy2 should be passed into the dense() on line 345.
Actual behavior
layer13_out is cloned on line 173 to layer71_cpy1 and layer71_cpy2.
layer71_cpy1 should be passed into the pooling1d_cl() on line 177 which works currently.
layer71_cpy2 should be passed into the dense() on line 345 but instead layer13_out is passed in which is incorrect.
A possible fix for the concatenate1d function bug is below.
Line 392 can be changed to: out_data[j + (i*input1_T::size)] = in_data1[j];Line 400 can be changed to: out_data[j + (i*input2_T::size) + (CONFIG_T::n_elem1_0)] = in_data2[j];
Quick summary
During the writing of an HLS project for a model with a skip connection, only one of the two copies of the stream are used in the first "branch". The original stream is propigated to the second "branch" instead of the second copy. This results in reading from a stream that is already empty. Commit hash: 107589fb1dac9aa898baea328acc6d519596e19d TF Model architecture
In addition tot this, the concatenate1d() function of nnet_merge_stream.h has a error in both loops where the out_data assignments don't have an additional offset to account for the progress of the outer loop. The fixed version is given below.
Details
The appropriate output stream is cloned with nnet::clone_stream<>() (layer13_out stream becomes layer71_cpy1 & layer71_cpy2). layer71_cpy1 is then correctly passed to the first "branch" but layer13_out (which has already been read from for the cloning process) is incorrectly passed to the second branch instead of layer71_cpy2. These warnings are produced as a result along with incorrect predictions.
Steps to Reproduce
Download & unzip files.zip
conda env create -f environment.ymlconda activate bepfm_fpga_tempgit clone https://github.com/fastmachinelearning/hls4ml.gitb.cd hls4ml;pip install .;cd ..python3 compile_prj.pyExpected behavior
layer13_out is cloned on line 173 of myproject.cpp to layer71_cpy1 and layer71_cpy2. layer71_cpy1 should be passed into the pooling1d_cl() on line 177 which works currently. layer71_cpy2 should be passed into the dense() on line 345.
Actual behavior
layer13_out is cloned on line 173 to layer71_cpy1 and layer71_cpy2. layer71_cpy1 should be passed into the pooling1d_cl() on line 177 which works currently. layer71_cpy2 should be passed into the dense() on line 345 but instead layer13_out is passed in which is incorrect.
files.zip
Possible fix
A possible fix for the concatenate1d function bug is below. Line 392 can be changed to:
out_data[j + (i*input1_T::size)] = in_data1[j];Line 400 can be changed to:out_data[j + (i*input2_T::size) + (CONFIG_T::n_elem1_0)] = in_data2[j];