// Propagation of tile backdrop for filling.
//
// Each thread reads one path element and calculates the number of spanned tiles
// based on the bounding box.
// In a further compaction step, the workgroup loops over the corresponding tile rows per element in parallel.
// For each row the per tile backdrop will be read, as calculated in the previous coarse path segment kernel,
// and propagated from the left to the right (prefix summed).
//
// Output state:
//  - Each path element has an array of tiles covering the whole path based on boundig box
//  - Each tile per path element contains the 'backdrop' and a list of subdivided path segments

#version 450
#extension GL_GOOGLE_include_directive : enable

#include "setup.h"

#define LG_BACKDROP_WG (7 + LG_WG_FACTOR)
#define BACKDROP_WG (1 << LG_BACKDROP_WG)

layout(local_size_x = BACKDROP_WG, local_size_y = 1) in;

layout(set = 0, binding = 0) buffer AnnotatedBuf {
    uint[] annotated;
};

// This is really only used for n_elements; maybe we can handle that
// a different way, but it's convenient to have the same signature as
// tile allocation.
layout(set = 0, binding = 1) readonly buffer AllocBuf {
    uint n_elements; // paths
    uint n_pathseg;
    uint alloc;
};

layout(set = 0, binding = 2) buffer TileBuf {
    uint[] tile;
};

#include "annotated.h"
#include "tile.h"

shared uint sh_row_count[BACKDROP_WG];
shared uint sh_row_base[BACKDROP_WG];
shared uint sh_row_width[BACKDROP_WG];

void main() {
    uint th_ix = gl_LocalInvocationID.x;
    uint element_ix = gl_GlobalInvocationID.x;
    AnnotatedRef ref = AnnotatedRef(element_ix * Annotated_size);

    // Work assignment: 1 thread : 1 path element
    uint row_count = 0;
    if (element_ix < n_elements) {
        uint tag = Annotated_tag(ref);
        if (tag == Annotated_Fill) {
            PathRef path_ref = PathRef(element_ix * Path_size);
            Path path = Path_read(path_ref);
            sh_row_width[th_ix] = path.bbox.z - path.bbox.x;
            row_count = path.bbox.w - path.bbox.y;
            if (row_count == 1) {
                // Note: this can probably be expanded to width = 2 as
                // long as it doesn't cross the left edge.
                row_count = 0;
            }
            sh_row_base[th_ix] = (path.tiles.offset >> 2) + 1;
        }
    }

    sh_row_count[th_ix] = row_count;
    // Prefix sum of sh_row_count
    for (uint i = 0; i < LG_BACKDROP_WG; i++) {
        barrier();
        if (th_ix >= (1 << i)) {
            row_count += sh_row_count[th_ix - (1 << i)];
        }
        barrier();
        sh_row_count[th_ix] = row_count;
    }
    barrier();
    // Work assignment: 1 thread : 1 path element row
    uint total_rows = sh_row_count[BACKDROP_WG - 1];
    for (uint row = th_ix; row < total_rows; row += BACKDROP_WG) {
        // Binary search to find element
        uint el_ix = 0;
        for (uint i = 0; i < LG_BACKDROP_WG; i++) {
            uint probe = el_ix + ((BACKDROP_WG / 2) >> i);
            if (row >= sh_row_count[probe - 1]) {
                el_ix = probe;
            }
        }
        uint seq_ix = row - (el_ix > 0 ? sh_row_count[el_ix - 1] : 0);
        uint width = sh_row_width[el_ix];
        // Process one row sequentially
        // Read backdrop value per tile and prefix sum it
        uint tile_el_ix = sh_row_base[el_ix] + seq_ix * 2 * width;
        uint sum = tile[tile_el_ix];
        for (uint x = 1; x < width; x++) {
            tile_el_ix += 2;
            sum += tile[tile_el_ix];
            tile[tile_el_ix] = sum;
        }
    }
}