#include "gpu_offset.h"
#include <stdio.h>
#include <stdlib.h>
#include <math.h>

int main() {
    printf("=== C API Example: Linking libcudaslicer_offset.so ===\n");

    // 1. Initialize GPU
    if (gpu_offset_init(0) != 0) {
        fprintf(stderr, "Failed to initialize GPU\n");
        return 1;
    }

    // 2. Create a test slice layer with a square polygon and a circular hole
    GpuContour contours[2];

    // Square exterior: 20mm x 20mm centered at origin
    GpuPoint2D square_pts[4] = {
        { -10.0f, -10.0f },
        {  10.0f, -10.0f },
        {  10.0f,  10.0f },
        { -10.0f,  10.0f }
    };
    contours[0].points = square_pts;
    contours[0].num_points = 4;
    contours[0].is_hole = 0;

    // Diamond interior hole
    GpuPoint2D hole_pts[4] = {
        {  0.0f, -4.0f },
        {  4.0f,  0.0f },
        {  0.0f,  4.0f },
        { -4.0f,  0.0f }
    };
    contours[1].points = hole_pts;
    contours[1].num_points = 4;
    contours[1].is_hole = 1;

    GpuSliceLayer layer;
    layer.z_height = 0.20f;
    layer.contours = contours;
    layer.num_contours = 2;

    // 3. Configure Offsets
    float offset_dists[2] = { 0.45f, 0.90f };
    GpuOffsetConfig config;
    config.pitch_mm = 0.05f;
    config.pad_mm = 1.0f;
    config.offset_distances_mm = offset_dists;
    config.num_offsets = 2;
    config.use_gpu_rasterizer = 1;

    // 4. Run GPU Offset
    GpuLayerResult* results = NULL;
    GpuOffsetStats stats = {0};

    int ret = gpu_offset_polygon_layers(
        &layer, 1,
        -10.0f, -10.0f, 10.0f, 10.0f,
        &config, &results, &stats
    );

    if (ret == 0 && results) {
        printf("GPU Offset Successful!\n");
        printf("  Rasterization:   %.3f ms\n", stats.t_raster_ms);
        printf("  H2D Transfer:    %.3f ms\n", stats.t_h2d_ms);
        printf("  GPU JFA Compute: %.3f ms\n", stats.t_gpu_jfa_ms);
        printf("  D2H Transfer:    %.3f ms\n", stats.t_d2h_ms);
        printf("  Vectorization:   %.3f ms\n", stats.t_vectorize_ms);
        printf("  Total Time:      %.3f ms\n", stats.t_total_pipeline_ms);

        for (size_t o = 0; o < results[0].num_perimeters; ++o) {
            printf("  Perimeter %zu (Offset -%.2f mm): %zu vector loops\n",
                   o + 1, results[0].perimeters[o].offset_distance_mm,
                   results[0].perimeters[o].num_contours);
        }

        gpu_offset_free_results(results, 1);
    } else {
        fprintf(stderr, "GPU Offset failed!\n");
    }

    gpu_offset_cleanup();
    printf("=== Done ===\n");
    return 0;
}