// clang -Oz --target=wasm32 -nostdlib -Wl,--export-all,--no-entry -msimd128 lod.c -o lod.wasm

#include <wasm_simd128.h>

extern char __heap_base;

static int allocated_static;
static int allocated_dynamic;

void
free_static(void)
{
    allocated_static = 0;
}

void
free_dynamic(void)
{
    allocated_dynamic = 0;
}

void *
alloc_static(int size)
{
    void *result = &__heap_base + allocated_static;
    allocated_static += size;
    return(result);
}

void *
alloc_dynamic(int size)
{
    void *result = &__heap_base + allocated_static + allocated_dynamic;
    allocated_dynamic += size;
    return(result);
}

static int
rdp_find_max(float *xs, float *ys, unsigned char *pressures, float zoom, int coords_from, 
             int segment_start, int segment_end)
{
    int result = -1;

    if (segment_start == segment_end) {
        return(result);
    }

    float EPS = 0.125f / zoom * 255.0f;
    float max_dist = 0.0f;

    float ax = xs[coords_from + segment_start];
    float ay = ys[coords_from + segment_start];
    float bx = xs[coords_from + segment_end];
    float by = ys[coords_from + segment_end];

    unsigned char ap = pressures[coords_from / 2 + segment_start];
    unsigned char bp = pressures[coords_from / 2 + segment_end];

    float dx = bx - ax;
    float dy = by - ay;
    
    float dist_ab = __builtin_sqrtf(dx * dx + dy * dy);
    float dir_nx = dy / dist_ab * 255.0f;
    float dir_ny = -dx / dist_ab * 255.0f;
#if 1
    for (int i = segment_start + 1; i < segment_end; ++i) {
        float px = xs[coords_from + i];
        float py = ys[coords_from + i];
        
        unsigned char pp = pressures[coords_from + i];

        float apx = px - ax;
        float apy = py - ay;
        
        float dist = __builtin_fabsf(apx * dir_nx + apy * dir_ny)
            + __builtin_abs(pp - ap) + __builtin_abs(pp - bp);
        
        if (dist > EPS && dist > max_dist) {
            result = i;
            max_dist = dist;
        }
    }
#else
    v128_t eps_x4 = wasm_f32x4_splat(EPS);
    v128_t ax_x4 = wasm_f32x4_splat(ax);
    v128_t ay_x4 = wasm_f32x4_splat(ay);
    v128_t ap_x4 = wasm_f32x4_splat(ap);
    v128_t bp_x4 = wasm_f32x4_splat(bp);
    v128_t dir_nx_x4 = wasm_f32x4_splat(dir_nx);
    v128_t dir_ny_x4 = wasm_f32x4_splat(dir_ny);

    v128_t index_x4 = wasm_u32x4_make(segment_start + 1, segment_start + 2, segment_start + 3, segment_start + 4);
    v128_t four_x4 = wasm_u32x4_const_splat(4);
    v128_t max_vals_x4 = wasm_f32x4_const_splat(0.0f);
    v128_t max_index_x4 = wasm_u32x4_const_splat(-1);

    for (int i = segment_start + 1; i < segment_end - 3; i += 4) {
        v128_t px_x4 = wasm_v128_load(xs + coords_from + i);
        v128_t py_x4 = wasm_v128_load(ys + coords_from + i);
        v128_t pp_x16 = wasm_v128_load(pressures + coords_from / 2 + i);

        // Take 4 highest bytes and convert to float
        v128_t pp_x8 = wasm_u16x8_extend_high_u8x16(pp_x16);
        v128_t pp_x4i = wasm_u32x4_extend_high_u16x8(pp_x8);
        v128_t pp_x4 = wasm_f32x4_convert_i32x4(pp_x4i); // i version is 8 times faster on x64?

        v128_t apx_x4 = wasm_f32x4_sub(px_x4, ax_x4);
        v128_t apy_x4 = wasm_f32x4_sub(py_x4, ay_x4);

        v128_t dist_x4 = wasm_f32x4_add(
            wasm_f32x4_add(
                wasm_f32x4_abs(wasm_f32x4_sub(pp_x4, ap_x4)),
                wasm_f32x4_abs(wasm_f32x4_sub(pp_x4, bp_x4))
            ),
            wasm_f32x4_abs(
                wasm_f32x4_add(
                    wasm_f32x4_mul(apx_x4, dir_nx_x4),
                    wasm_f32x4_mul(apy_x4, dir_ny_x4)
                )
            )
        );
        
        v128_t dist_mask = wasm_f32x4_gt(dist_x4, eps_x4);
        v128_t max_mask = wasm_f32x4_gt(dist_x4, max_vals_x4);
        v128_t mask = wasm_v128_and(dist_mask, max_mask);

        max_index_x4 = wasm_v128_or(
            wasm_v128_and(index_x4, mask),
            wasm_v128_andnot(max_index_x4, mask)
        );

        max_vals_x4 = wasm_v128_or(
            wasm_v128_and(dist_x4, mask),
            wasm_v128_andnot(max_vals_x4, mask)
        );
        
        index_x4 = wasm_i32x4_add(index_x4, four_x4);
    }

    int indices[4];
    float values[4];

    wasm_v128_store(indices, max_index_x4);
    wasm_v128_store(values, max_vals_x4);

    for (int i = 0; i < 4; ++i) {
        if (indices[i] != -1) {
            if (values[i] > max_dist) {
                result = indices[i];
                max_dist = values[i];
            }
        }
    }

    int remainder = (segment_end - segment_start - 1) % 4;

    for (int i = segment_end - remainder; i < segment_end; ++i) {
        float px = xs[coords_from + i];
        float py = ys[coords_from + i];

        unsigned char pp = pressures[coords_from + i];

        float apx = px - ax;
        float apy = py - ay;

        float dist = __builtin_fabsf(apx * dir_nx + apy * dir_ny)
            + __builtin_abs(pp - ap) + __builtin_abs(pp - bp);

        if (dist > EPS && dist > max_dist) {
            result = i;
            max_dist = dist;
        }
    }

#endif
    
    return(result);
}

int
do_lod(int *clipped_indices, int clipped_count, float zoom, 
       int *stroke_coords_from,
       float *line_threshold, 
       float *xs,
       float *ys,
       unsigned char *pressures,
       int coordinates_count)
{
    if (clipped_count == 0) {
        return(0);
    }

    int *segments_from = alloc_dynamic((clipped_count + 1) * 4);
    int *segments = alloc_dynamic(coordinates_count * 4);
    
    int segments_head = 0;
    int stack[4096]; // TODO: what's a reasonable max size for this?
    
    for (int i = 0; i < clipped_count; ++i) {
        int stroke_index = clipped_indices[i];
        
        // TODO: convert to a proper CSR, save half the memory
        int coords_from = stroke_coords_from[stroke_index];
        int coords_to = stroke_coords_from[stroke_index + 1];
        
        int point_count = coords_to - coords_from;
        
        // Basic CSR crap
        segments_from[i] = segments_head;
        
        if (zoom < line_threshold[stroke_index]) {
            // Fast paths for collapsing to a single line segment
            segments[segments_head++] = 0;
            segments[segments_head++] = point_count - 1;
            continue;
        }
        
        int segment_count = 2;
        int stack_head = 0;
        
        segments[segments_head++] = 0;
        
        stack[stack_head++] = 0;
        stack[stack_head++] = 0;
        stack[stack_head++] = point_count - 1;
        
        while (stack_head > 0) {
            int end = stack[--stack_head];
            int start = stack[--stack_head];
            int type = stack[--stack_head];
            
            if (type == 1) {
                segments[segments_head++] = start;
            } else {
                int max = rdp_find_max(xs, ys, pressures, zoom, coords_from, start, end);
                if (max != -1) {
                    segment_count += 1;
                    
                    stack[stack_head++] = 0;
                    stack[stack_head++] = max;
                    stack[stack_head++] = end;
                    
                    stack[stack_head++] = 1;
                    stack[stack_head++] = max;
                    stack[stack_head++] = -1;
                    
                    stack[stack_head++] = 0;
                    stack[stack_head++] = start;
                    stack[stack_head++] = max;
                }
            }
        }
        
        segments[segments_head++] = point_count - 1;
        
        if (segment_count == 2 && zoom > line_threshold[stroke_index]) {
            line_threshold[stroke_index] = zoom;
        }
    }

    segments_from[clipped_count] = segments_head;

    // Write actual coordinates (points) and stroke ids
    float *points = alloc_dynamic(segments_head * 2 * 4);
    int *ids = alloc_dynamic(segments_head * 4);
    unsigned char *pressures_res = alloc_dynamic(segments_head);

    int phead = 0;
    int ihead = 0;

    for (int i = 0; i < clipped_count; ++i) {
        int stroke_index = clipped_indices[i];
        
        // TODO: convert to a proper CSR, save half the memory
        int base_stroke = stroke_coords_from[stroke_index];
        int from = segments_from[i];
        int to = segments_from[i + 1];

        for (int j = from; j < to; ++j) {
            int point_index = segments[j];
            float x = xs[base_stroke + point_index];
            float y = ys[base_stroke + point_index];

            points[phead++] = x;
            points[phead++] = y;

            pressures_res[ihead] = pressures[base_stroke + point_index];

            if (j != to - 1) {
                ids[ihead++] = stroke_index;
            } else {
                ids[ihead++] = stroke_index | (1 << 31);
            }
        }
    }
    
    return(segments_head);
}