// SPDX-License-Identifier: Apache-2.0 OR MIT OR Unlicense

// The reduction phase for prefix sum implemented as a tree reduction.

#version 450

#define N_ROWS 8
#define LG_WG_SIZE 9
#define WG_SIZE (1 << LG_WG_SIZE)
#define PARTITION_SIZE (WG_SIZE * N_ROWS)

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

struct Monoid {
    uint element;
};

layout(set = 0, binding = 0) readonly buffer InBuf {
    Monoid[] inbuf;
};

layout(set = 0, binding = 1) buffer OutBuf {
    Monoid[] outbuf;
};

shared Monoid sh_scratch[WG_SIZE];

Monoid combine_monoid(Monoid a, Monoid b) {
    return Monoid(a.element + b.element);
}

void main() {
    uint ix = gl_GlobalInvocationID.x * N_ROWS;
    // TODO: gate buffer read
    Monoid agg = inbuf[ix];
    for (uint i = 1; i < N_ROWS; i++) {
        agg = combine_monoid(agg, inbuf[ix + i]);
    }
    sh_scratch[gl_LocalInvocationID.x] = agg;
    for (uint i = 0; i < LG_WG_SIZE; i++) {
        barrier();
        // We could make this predicate tighter, but would it help?
        if (gl_LocalInvocationID.x + (1u << i) < WG_SIZE) {
            Monoid other = sh_scratch[gl_LocalInvocationID.x + (1u << i)];
            agg = combine_monoid(agg, other);
        }
        barrier();
        sh_scratch[gl_LocalInvocationID.x] = agg;
    }
    if (gl_LocalInvocationID.x == 0) {
        outbuf[gl_WorkGroupID.x] = agg;
    }
}