// Mechanical adaptation of OA's keep-dimension Sum reduction.
import storage;
import attributes;
struct PushConstants {
uint input_index;
uint output_index;
uint outer_size;
uint dim_size;
uint inner_size;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("sum_axis")]
[domain("matrix")]
[variant("generic")]
[dtype("f32")]
[status("experimental")]
[shader("compute")]
[numthreads(256, 1, 1)]
void main(uint3 dispatch_thread_id : SV_DispatchThreadID) {
uint output_index = dispatch_thread_id.x;
uint output_count = push.outer_size * push.inner_size;
if (output_index >= output_count) {
return;
}
uint inner = output_index % push.inner_size;
uint outer = output_index / push.inner_size;
uint base = outer * push.dim_size * push.inner_size + inner;
float accumulator = 0.0f;
for (uint axis_index = 0; axis_index < push.dim_size; ++axis_index) {
accumulator += load_f32(
storage_buffers[push.input_index],
base + axis_index * push.inner_size
);
}
store_f32(storage_buffers[push.output_index], output_index, accumulator);
}