import storage;
import attributes;
struct PushConstants {
uint input_index;
uint output_index;
uint channels;
uint values_per_channel;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("mel_normalize")]
[domain("audio")]
[variant("generic")]
[dtype("f32")]
[status("experimental")]
[shader("compute")]
[numthreads(1, 1, 1)]
void main(uint3 thread_id : SV_DispatchThreadID) {
uint channel = thread_id.x;
if (channel >= push.channels) return;
uint base = channel * push.values_per_channel;
float mean = 0.0f;
for (uint index = 0; index < push.values_per_channel; ++index) {
mean += load_f32(storage_buffers[push.input_index], base + index);
}
mean /= float(push.values_per_channel);
float variance = 0.0f;
for (uint index = 0; index < push.values_per_channel; ++index) {
float centered = load_f32(storage_buffers[push.input_index], base + index) - mean;
variance += centered * centered;
}
float inverse_stddev = rsqrt(variance / float(push.values_per_channel) + 1e-5f);
for (uint index = 0; index < push.values_per_channel; ++index) {
float centered = load_f32(storage_buffers[push.input_index], base + index) - mean;
store_f32(storage_buffers[push.output_index], base + index, centered * inverse_stddev);
}
}