import attributes;
import storage;
struct PushConstants {
uint input_index;
uint indices_index;
uint output_index;
uint rows;
uint input_width;
uint selected_width;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("gather_last_dim")]
[domain("matrix")]
[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.rows * push.selected_width) {
return;
}
uint row = index / push.selected_width;
int selected = load_i32(storage_buffers[push.indices_index], index);
float value = selected >= 0 && uint(selected) < push.input_width
? load_f32(storage_buffers[push.input_index], row * push.input_width + uint(selected))
: 0.0;
store_f32(storage_buffers[push.output_index], index, value);
}