import attributes;
import storage;
struct PushConstants {
uint gradient_output_index;
uint indices_index;
uint gradient_input_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_backward")]
[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.input_width) {
return;
}
uint row = index / push.input_width;
uint column = index % push.input_width;
float sum = 0.0;
for (uint selected = 0; selected < push.selected_width; ++selected) {
uint route = row * push.selected_width + selected;
int source = load_i32(storage_buffers[push.indices_index], route);
if (source >= 0 && uint(source) == column) {
sum += load_f32(storage_buffers[push.gradient_output_index], route);
}
}
store_f32(storage_buffers[push.gradient_input_index], index, sum);
}