/* Phase 2b: RMSNorm in C, validated against Python ref.
 * RMSNorm(x)[i] = x[i] * rsqrt(mean(x^2)+eps) * weight[i]   (per row, over dim=6144)
 * Build: gcc -O2 -o phase2b_rmsnorm phase2b_rmsnorm.c -lm
 */
#include <stdio.h>
#include <stdlib.h>
#include <math.h>

#define DIM 6144
#define ROWS 5
#define EPS 1e-5f

static float* load_f32(const char* path, long n) {
    FILE* f = fopen(path, "rb");
    if (!f) { perror(path); exit(1); }
    float* buf = malloc(n * sizeof(float));
    if (fread(buf, sizeof(float), n, f) != (size_t)n) { fprintf(stderr, "short read %s\n", path); exit(1); }
    fclose(f); return buf;
}

int main(void) {
    const char* D = "ref";
    char p[512];
    snprintf(p, sizeof p, "%s/norm_in.f32", D);     float* x = load_f32(p, ROWS * DIM);
    snprintf(p, sizeof p, "%s/norm_weight.f32", D); float* w = load_f32(p, DIM);
    snprintf(p, sizeof p, "%s/norm_out.f32", D);    float* ref = load_f32(p, ROWS * DIM);

    float* out = malloc(ROWS * DIM * sizeof(float));
    for (int r = 0; r < ROWS; r++) {
        const float* xr = x + r * DIM;
        double ss = 0.0;
        for (int i = 0; i < DIM; i++) ss += (double)xr[i] * xr[i];
        float inv = 1.0f / sqrtf((float)(ss / DIM) + EPS);
        for (int i = 0; i < DIM; i++) out[r * DIM + i] = xr[i] * inv * w[i];
    }

    /* validate against Python ref */
    double max_abs = 0.0; long worst = 0;
    for (long i = 0; i < (long)ROWS * DIM; i++) {
        double d = fabs((double)out[i] - ref[i]);
        if (d > max_abs) { max_abs = d; worst = i; }
    }
    printf("C RMSNorm out[0..8]: ");
    for (int i = 0; i < 8; i++) printf("%.4f ", out[i]);
    printf("\nmax abs diff vs Python ref: %.3e (at idx %ld)\n", max_abs, worst);
    printf("VERDICT: %s\n", max_abs < 1e-4 ? "PASS — RMSNorm matches Python" : "FAIL");
    return max_abs < 1e-4 ? 0 : 1;
}