/*
 * nnrun.c - run the quantised MLP over a test set and measure it (C89).
 *
 * Reads a test file written by train.py:
 *   uint32 count (little-endian), then count*784 pixel bytes, then count labels.
 *
 * Usage: nnrun [testfile] [max-samples]
 *
 * DeepSeek (deepseek-v4.1-flash), October 2026.
 */
#include <stdio.h>
#include <stdlib.h>
#include <time.h>
#include "nn.h"

int main(int argc, char **argv)
{
    const char *path = "nn_test.bin";
    FILE *f;
    unsigned char hdr[4];
    unsigned char *pix, *lab;
    unsigned int n, i, limit;
    long correct = 0;
    clock_t t0, t1;
    double secs;

    if (argc > 1)
        path = argv[1];

    f = fopen(path, "rb");
    if (f == NULL) {
        perror(path);
        return 1;
    }
    if (fread(hdr, 1, 4, f) != 4) {
        fprintf(stderr, "short header\n");
        return 1;
    }
    n = (unsigned) hdr[0] | ((unsigned) hdr[1] << 8) |
        ((unsigned) hdr[2] << 16) | ((unsigned) hdr[3] << 24);

    pix = (unsigned char *) malloc((size_t) n * 784);
    lab = (unsigned char *) malloc((size_t) n);
    if (pix == NULL || lab == NULL) {
        fprintf(stderr, "out of memory\n");
        return 1;
    }
    if (fread(pix, 1, (size_t) n * 784, f) != (size_t) n * 784 ||
        fread(lab, 1, (size_t) n, f) != (size_t) n) {
        fprintf(stderr, "short data\n");
        return 1;
    }
    fclose(f);

    limit = n;
    if (argc > 2) {
        unsigned int l = (unsigned int) atoi(argv[2]);
        if (l < n)
            limit = l;
    }

    t0 = clock();
    for (i = 0; i < limit; i++)
        if (nn_classify(pix + (size_t) i * 784) == lab[i])
            correct++;
    t1 = clock();

    secs = (double) (t1 - t0) / (double) CLOCKS_PER_SEC;
    if (secs <= 0.0)
        secs = 1e-9;

    printf("samples    %u\n", limit);
    printf("correct    %ld\n", correct);
    printf("accuracy   %.2f%%\n", 100.0 * (double) correct / (double) limit);
    printf("time       %.3f s\n", secs);
    printf("per infer  %.1f us\n", secs * 1e6 / (double) limit);
    printf("rate       %.0f inferences/s\n", (double) limit / secs);
    return 0;
}