// Two-input SwiGLU adjoint adapted from OA's donor kernel.
// Donor: source/cpp/lib/oa/ml/shader/compute/nn/swiglu/swigluBwd.slang
import activations;
import storage;
import attributes;
struct PushConstants {
uint gate_index;
uint up_index;
uint output_gradient_index;
uint gate_gradient_index;
uint up_gradient_index;
uint element_count;
};
[[vk::push_constant]] PushConstants push;
[[vk::binding(0, 0)]] RWByteAddressBuffer storage_buffers[];
[kernel_name("swiglu_backward")]
[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 gate = load_f32(storage_buffers[push.gate_index], index);
float up = load_f32(storage_buffers[push.up_index], index);
float output_gradient = load_f32(storage_buffers[push.output_gradient_index], index);
store_f32(
storage_buffers[push.gate_gradient_index],
index,
output_gradient * up * silu_grad(gate)
);
store_f32(
storage_buffers[push.up_gradient_index],
index,
output_gradient * silu(gate)
);
}