AWS: getting around no AXI stream support

[cholan2100]

Any one had any luck with using HLS generated with hls4ml on AWS platform? AWS platform lacks direct AXI stream support in its shell. So have to rely on copying memory buffers. CNN on hls4ml needs io_stream.

I tried the trick of modifying HLS kernel to have a wrapper around the real kernel, to push into hls:stream from memory buffers. It works alright on Csim, but not on cosim or hw, kernel seem stuck.

#include "myproject.h"

extern "C" {
    void hls4ml_kernel(input_t *input, result_t *output) {
        // create hls streams for input and output
        hls::stream<input_t> input_stream("input_stream");
#define DATA_SIZE_IN  (N_INPUT_1_1*N_INPUT_2_1*N_INPUT_3_1)
        for (unsigned int i = 0; i < DATA_SIZE_IN; i++) {
// #pragma HLS PIPELINE
// #pragma HLS UNROLL factor=1
            input_stream.write(input[i]);
        }

        hls::stream<result_t> output_stream("output_stream");
        myproject(input_stream, output_stream);

#define DATA_SIZE_OUT (1) // all 10 classes in one array
        for (unsigned int i = 0; i < DATA_SIZE_OUT; i++) {
            output[i] = output_stream.read();
        }
    }
}

any clues.?

ps: i have ported hand written implementation for mnist dataset classification to AWS. Would be nice to get hls4ml work with AWS for other real projects, as backend=VivadoAccelerator still doesn't support AWS platform. aws_cnn_digidet