// Mechanical adaptation of OA's per-element absolute-error kernel.
import storage;
import attributes;
struct PushConstants {
uint prediction_index;
uint target_index;
uint output_index;
uint element_count;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("l1")]
[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 prediction = load_f32(storage_buffers[push.prediction_index], index);
float target = load_f32(storage_buffers[push.target_index], index);
store_f32(storage_buffers[push.output_index], index, abs(prediction - target));
}