// SGD-owned in-place FP32 update with optional coupled weight decay.
import storage;
import attributes;
struct PushConstants {
uint parameter_index;
uint gradient_index;
uint element_count;
float learning_rate;
float weight_decay;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("sgd")]
[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.element_count) {
return;
}
float parameter = load_f32(storage_buffers[push.parameter_index], index);
float gradient = load_f32(storage_buffers[push.gradient_index], index);
if (push.weight_decay > 0.0f) {
gradient += push.weight_decay * parameter;
}
parameter -= push.learning_rate * gradient;
store_f32(storage_buffers[push.parameter_index], index, parameter);
}