#include "op_plugin/OpApiInterface.h"
#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "torch_npu/csrc/framework/utils/InternalFormatOpAdapter.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
at::Tensor gelu_backward(const at::Tensor & grad_output, const at::Tensor & self, c10::string_view approximate)
{
static auto gelu_sc = at_npu::native::env::CheckCompatibleImpl();
auto output_size_0 = op_infer::broadcast_ops_npu_output_size(grad_output, self);
auto output_dtype_0 = at::native::result_type(grad_output, self);
at::Tensor grad_input = npu_preparation::apply_tensor_without_format(output_size_0,
grad_output.options().dtype(output_dtype_0));
auto approximate_str = op_infer::npu_gelu_approximate_str(approximate);
auto approximate_ptr = const_cast<char *>(approximate_str.c_str());
if (!gelu_sc) {
DO_COMPATIBILITY(aclnnGeluBackward, acl_op::gelu_backward(grad_output, self, approximate));
EXEC_NPU_CMD(aclnnGeluBackward, grad_output, self, grad_input);
} else {
EXEC_NPU_CMD(aclnnGeluBackwardV2, grad_output, self, approximate_ptr, grad_input);
}
return grad_input;
}
}