[[vk::binding(0, 0)]]
StructuredBuffer<float2> input;

[[vk::binding(1, 0)]]
RWStructuredBuffer<float2> output;

#define PI 3.1415926535897932384626433832795028841971f
#define TAU 6.283185307179586476925286766559005768394f

[shader("compute")]
[numthreads(256, 1, 1)]
void brute_force_dft(uint3 globalId : SV_DispatchThreadID)
{

    uint input_len, input_stride;
    input.GetDimensions(input_len, input_stride);

    const float k = globalId.x;
    const float N = input_len;

    float phase_base = -TAU * k / N;
    float2 result = float2(0, 0);
    for (uint n = 0; n < input_len; n++)
    {
        float phase = phase_base * n;
        float re, im;
        sincos(phase, im, re);
        float2 x_n = input[n];
        result += float2(x_n[0] * re - x_n[1] * im, x_n[0] * im + x_n[1] * re);
    }

    output[globalId.x] = result;
}