import attributes;
import philox;
import storage;
struct PushConstants {
uint input_index;
uint output_index;
uint element_count;
uint seed_low;
uint seed_high;
uint offset;
float mean;
float stddev;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("philox_normal")]
[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 * 2u;
if (base >= push.element_count) {
return;
}
uint4 bits = philox_rng(
uint4(thread + push.offset, 0u, 0u, 0u),
uint2(push.seed_low, push.seed_high)
);
float first;
float second;
philox_to_normal(bits, first, second);
store_f32(storage_buffers[push.output_index], base, first * push.stddev + push.mean);
if (base + 1u < push.element_count) {
store_f32(storage_buffers[push.output_index], base + 1u, second * push.stddev + push.mean);
}
}