Modifying output ports of generated IP

[Rezef]

Hello, is it possible to modify the ports of the generated IP based on the strategy. For example if you want to have an output port which is scalar instead of a vector or an hls::stream using python script? I am utilzing extension api to generate a corresponding scalar value but still having to stream the output by copying the scalar for all the channels which seems wasteful

[bo3z]

Not directly from Python, no. But you can modify the generated code from hls4ml manually in your output directory in the myproject.cpp/.hpp before running HLS synthesis etc. from the comand line using vivado_hls -f build_prj.tcl or vitis_hls -f build_prj.tcl

[Rezef]

So I need to configure the files of core_template and nnet_mult.h within the template folder of hls4ml so that the changes can be applied for all multiplications?

[bo3z]

Yes. And also, if you need to introduce any new data types that are applicable to your new multiplication.

[Rezef]

I have another question. Why are convolutions implemented as matrix multiplication, won't the resource usage be massive if we try to implement the entire feature map as a matrix multiplication supposedly if we have 200x200x32 then the patch matrix will be of size almost 39204x288 or similar. Even if we use grouping meaning computing matrix multiplication in chunks, the DSPs remain inactive throughout this storage. On the other hand, if we utilize normal dot product between the feature map extracted window and the kernel the dsps remain active throughout. Can u tell me why are convolutions considered slower compared to matrix multiplication?

[vloncar]

We don't do matrix-matrix multiplication, we implement convolution as a series of matrix-vector multiplications, the full patch matrix is never fully created/stored on the FPGA. In each iteration we prepare one vector of it and do the multiplication. We can then achieve speedup by grouping several iterations of that loop and execute them in parallel, while pipelining the loop so DSPs do remain active. We call these groups "partitions" in the code.

[Rezef]

Thanks for the reply. I was confused regarding the partition first and thought you are just partitioning the access of the feature map in groups rather than partitioning the channel dimension