// U32 Embedding gather. One invocation writes one output scalar.
import storage;
import attributes;
struct PushConstants {
uint weight_index;
uint indices_index;
uint output_index;
uint index_count;
uint num_embeddings;
uint embedding_dim;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("embedding")]
[domain("ml")]
[variant("generic")]
[dtype("f32")]
[status("experimental")]
[shader("compute")]
[numthreads(256, 1, 1)]
void main(uint3 dispatch_thread_id : SV_DispatchThreadID) {
uint output_index = dispatch_thread_id.x;
uint output_count = push.index_count * push.embedding_dim;
if (output_index >= output_count) {
return;
}
uint index_position = output_index / push.embedding_dim;
uint feature = output_index % push.embedding_dim;
uint row = storage_buffers[push.indices_index].Load(index_position * 4u);
float value = row < push.num_embeddings
? load_f32(storage_buffers[push.weight_index], row * push.embedding_dim + feature)
: asfloat(0x7fc00000u);
store_f32(storage_buffers[push.output_index], output_index, value);
}