// Donor: source/cpp/lib/oa/ml/shader/compute/ops/gather.slang
// FP32 specialization for stable expert-major MoE packing.
import storage;
import attributes;
struct PushConstants {
uint input_index;
uint packed_token_index;
uint output_index;
uint routes;
uint row_width;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("moe_gather")]
[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.routes * push.row_width) return;
uint packed_row = index / push.row_width;
uint column = index % push.row_width;
uint token = storage_buffers[push.packed_token_index].Load(packed_row * 4);
store_f32(
storage_buffers[push.output_index],
index,
load_f32(storage_buffers[push.input_index], token * push.row_width + column));
}