#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor>npu_add_layer_norm_backward(const c10::optional<at::Tensor> &dy_opt,
const at::Tensor &x1,
const at::Tensor &x2,
const at::Tensor &rstd,
const at::Tensor &mean,
const at::Tensor &gamma,
const c10::optional<at::Tensor> &dsum_opt)
{
DO_COMPATIBILITY(aclnnAddLayerNormGrad, acl_op::npu_add_layer_norm_backward(dy_opt, x1, x2, rstd, mean, gamma, dsum_opt));
const at::Tensor& dy = c10::value_or_else(dy_opt, [] {return at::Tensor();});
const at::Tensor& dsum = c10::value_or_else(dsum_opt, [] {return at::Tensor();});
auto d_y = dy.defined() ? dy : at::zeros(x1.sizes(), x1.options());
auto d_sum = dsum.defined() ? dsum : at::zeros(x1.sizes(), x1.options());
at::SmallVector<int64_t, SIZE> shape;
for (int64_t index = 0; index < gamma.dim(); index++) {
shape.emplace_back(gamma.size(index));
}
at::Tensor dx = npu_preparation::apply_tensor(x1);
at::Tensor dgamma = npu_preparation::apply_tensor(shape, x1.options().dtype(at::kFloat), x1);
at::Tensor dbeta = npu_preparation::apply_tensor(shape, x1.options().dtype(at::kFloat), x1);
EXEC_NPU_CMD(aclnnAddLayerNormGrad, d_y, x1, x2, rstd, mean, gamma, d_sum, dx, dgamma, dbeta);
return std::make_tuple(dx, dx, dgamma, dbeta);
}
}