// Adapted from OA C++ rlRolloutAppend.slang and matrixCopyRegion.slang.
// Each invocation owns one observation element and, for index<count, one
// aligned scalar in all four statistic/action outputs. Actions remain I32.
import storage;
import attributes;
struct PushConstants {
uint observation_index;
uint action_index;
uint old_log_probability_index;
uint advantage_index;
uint returns_index;
uint output_observation_index;
uint output_action_index;
uint output_old_log_probability_index;
uint output_advantage_index;
uint output_returns_index;
uint start;
uint count;
uint observation_elements;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("minibatch")]
[domain("ml")]
[variant("categorical_contiguous")]
[dtype("f32")]
[status("experimental")]
[shader("compute")]
[numthreads(256, 1, 1)]
void main(uint3 tid : SV_DispatchThreadID) {
uint i = tid.x;
if (i < push.count * push.observation_elements) {
store_f32(storage_buffers[push.output_observation_index], i,
load_f32(storage_buffers[push.observation_index], push.start * push.observation_elements + i));
}
if (i >= push.count) return;
uint source = push.start + i;
store_i32(storage_buffers[push.output_action_index], i, load_i32(storage_buffers[push.action_index], source));
store_f32(storage_buffers[push.output_old_log_probability_index], i, load_f32(storage_buffers[push.old_log_probability_index], source));
store_f32(storage_buffers[push.output_advantage_index], i, load_f32(storage_buffers[push.advantage_index], source));
store_f32(storage_buffers[push.output_returns_index], i, load_f32(storage_buffers[push.returns_index], source));
}