#include "test_case.h"

int main() {
    // Top-level AXI4-Stream ports for the DUT
    hls::stream<coeff_axis_big_t> in_data;
    hls::stream<coeff_axis_t>     out_data;
    coeff_axis_big_t local_stream1;
    coeff_axis_t     local_stream2;
    coeff_t actual_outputs[Nt];
    int i;

    // Write input stimuli into input AXI4-Stream
    for (i = 0; i < Nt; i++) {
        coeff_t        val1 = input1_vals[i];
        double_coeff_t val2 = (double_coeff_t) input2_vals[i] << 24;
        // Pack two 24-bit values into one 48-bit word (stored in ap_axiu<48>)
        local_stream1.data = (ap_uint<48>) ((ap_uint<48>) val1 | (ap_uint<48>) val2);
        local_stream1.keep = -1;
        local_stream1.strb = -1;
        local_stream1.last = (i == Nt - 1) ? 1 : 0;
        in_data.write(local_stream1);
    }

    // Call DUT
    poly_mult(in_data, out_data);

    // Read results from output AXI4-Stream
    for (i = 0; i < Nt; i++) {
        local_stream2 = out_data.read();
        actual_outputs[i] = (coeff_t) local_stream2.data;
        // (Optionally check local_stream2.last here)
    }

    // Compare against golden output
    int ret_val = 0;
    for (i = 0; i < Nt; i++) {
        if (output_vals[i] != actual_outputs[i]) {
            ret_val++;
            std::cout << "Mismatch at index " << i 
                      << ": expected " << output_vals[i] 
                      << ", got " << actual_outputs[i] << std::endl;
            break;
        }
    }
    return ret_val;
}