#include <CL/sycl.hpp>
#include <iostream>

#define SIZE 10

using namespace cl::sycl;

int main() {
    float input[SIZE];
    float output[SIZE];

    // Initialize input array
    for (int i = 0; i < SIZE; i++) {
        input[i] = i + 1;
    }

    // Create SYCL queue
    queue q;

    // Submit command group for execution
    q.submit([&](handler& cgh) {
        // Create accessors for input and output buffers
        accessor inputAccessor(input, range<1>(SIZE), cgh);
        accessor outputAccessor(output, range<1>(SIZE), cgh);

        // Define the kernel
        cgh.parallel_for(range<1>(SIZE), [=](id<1> idx) {
            outputAccessor[idx] = inputAccessor[idx] * inputAccessor[idx];
        });
    });

    // Wait for the command group to finish
    q.wait();

    // Print the result
    for (int i = 0; i < SIZE; i++) {
        std::cout << output[i] << " ";
    }
    std::cout << std::endl;

    return 0;
}