import attributes;
import dropout_rng;
import philox;
import storage;
struct PushConstants {
uint input_index;
uint output_index;
uint state_index;
uint element_count;
uint seed_low;
uint seed_high;
uint offset;
float probability;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("dropout_replay")]
[domain("matrix")]
[variant("generic")]
[dtype("f32")]
[status("experimental")]
[shader("compute")]
[numthreads(256, 1, 1)]
void main(uint3 group_id : SV_GroupID, uint3 group_thread_id : SV_GroupThreadID) {
uint thread = group_id.x * 256u + group_thread_id.x;
uint base = thread * 4u;
if (base >= push.element_count) {
return;
}
uint replay = storage_buffers[push.state_index].Load(0u);
uint counters_per_replay = (push.element_count + 3u) / 4u;
uint4 bits = philox_rng(
uint4(thread + push.offset + replay * counters_per_replay, 0u, 0u, 0u),
uint2(push.seed_low, push.seed_high)
);
[unroll] for (uint lane = 0u; lane < 4u; ++lane) {
uint index = base + lane;
if (index < push.element_count) {
float input = load_f32(storage_buffers[push.input_index], index);
store_f32(storage_buffers[push.output_index], index, apply_dropout(input, bits[lane], push.probability));
}
}
}