import attributes;
import keccak;
struct PushConstants {
uint input_index;
uint output_index;
uint count;
uint hash_bytes;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("merkle_reduce")]
[domain("cryptography")]
[variant("generic")]
[dtype("u8")]
[status("experimental")]
[shader("compute")]
[numthreads(256, 1, 1)]
void main(uint3 dispatch_thread_id : SV_DispatchThreadID) {
uint index = dispatch_thread_id.x;
if (index >= push.count) return;
uint2 state[25];
for (uint lane = 0; lane < 25; ++lane) state[lane] = uint2(0, 0);
uint children_base = index * push.hash_bytes * 2;
for (uint lane = 0; lane < 8; ++lane) {
uint address = children_base + lane * 8;
state[lane] = uint2(
storage_buffers[push.input_index].Load<uint>(address),
storage_buffers[push.input_index].Load<uint>(address + 4)
);
}
state[8].x ^= 0x1fu;
state[16].y ^= 0x80000000u;
keccak_f1600_u32(state);
uint output_base = index * push.hash_bytes;
for (uint lane = 0; lane < 4; ++lane) {
uint address = output_base + lane * 8;
storage_buffers[push.output_index].Store<uint>(address, state[lane].x);
storage_buffers[push.output_index].Store<uint>(address + 4, state[lane].y);
}
}