alias number = f32;

struct VectorBatch {
  size: vec4<u32>,
  data: array<u32>,
}

struct MatrixBatch {
  size: vec4<u32>,
  data: array<vec4<number>>,
}

@group(0) @binding(0)
var<storage, read> v: VectorBatch;
@group(0) @binding(1)
var<storage, read> m: MatrixBatch;
@group(0) @binding(2)
var<storage, read_write> result: MatrixBatch;

@compute @workgroup_size(1)
fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
  result.size = m.size;

  var vc = result.size.x / 4;
  var resi = (global_id.z * result.size.y + global_id.y) * vc + global_id.x;
  result.data[resi] = vec4(0f);

  var mi = global_id.z * m.size.y * vc + global_id.x;
  var vi = global_id.z * v.size.x;
  for (var i = 0u; i < v.size.x; i++) {
    if (global_id.y == v.data[vi]) {
      result.data[resi] += m.data[mi];
    }
    mi += vc;
    vi++;
  }
}