verify.h

#pragma once
#include <cstddef>
#include <cstdint>
#include <cufft.h>
#include <vector>

namespace sbfft {

// Aggregate statistics over a complex N x N output. Computed on device.
struct OutputStats {
    double sum_sq_mag = 0; // SIGMA |Y[u,v]|^2 (Parseval check: equals N^2 * K)
    float max_mag = 0.f;
    double mean_mag = 0.0;
};

// Compute aggregate stats from a device-resident N x N complex array.
OutputStats compute_stats(const cufftComplex* d_out, size_t N);

// Compute max absolute element-wise difference between two host arrays of
// length len. Returns max |a[i] - b[i]| measured on the complex magnitude
// of the element difference (so it's robust to phase-only error reporting).
double max_abs_diff(const std::vector<cufftComplex>& a,
                    const std::vector<cufftComplex>& b);

// Parseval's identity for a binary input matrix with K ones produces:
//   SIGMA_uv |Y[u,v]|^2 = N^2 * K
// for the un-normalized forward DFT. Returns the relative error.
double parseval_relative_error(const OutputStats& stats,
                               int N, int64_t nnz);

} // namespace sbfft