// Donor: source/cpp/lib/oa/ml/shader/compute/nn/moe/moeGatherBwd.slang
// One invocation owns one output scalar and sums routes in stable slot order.
import storage;
import attributes;
struct PushConstants {
uint source_index;
uint inverse_index;
uint output_index;
uint routes;
uint row_width;
uint tokens;
uint routes_per_token;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("moe_gather_backward")]
[domain("ml")]
[variant("generic")]
[dtype("f32")]
[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.tokens * push.row_width) return;
uint token = index / push.row_width;
uint column = index % push.row_width;
float sum = 0.0f;
for (uint route_slot = 0; route_slot < push.routes_per_token; ++route_slot) {
uint packed = storage_buffers[push.inverse_index].Load(
(token * push.routes_per_token + route_slot) * 4);
if (packed < push.routes) {
sum += load_f32(
storage_buffers[push.source_index], packed * push.row_width + column);
}
}
store_f32(storage_buffers[push.output_index], index, sum);
}