#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/custom_functions/atb/AtbCommon.h"
#include <acl/acl.h>
using namespace std;
namespace atb {
using GroupTopkParam = atb::infer::GroupTopkParam;
void _npu_group_topk(const at::Tensor &self, int64_t k, int64_t group_num, int64_t n)
{
const c10::OptionalDeviceGuard device_guard(device_of(self));
OpParamCache<GroupTopkParam>& GroupTopkParamCache = OpParamCache<GroupTopkParam>::getInstance();
GroupTopkParam GroupTopkParam;
GroupTopkParam.groupNum = static_cast<int32_t>(group_num);
GroupTopkParam.k = static_cast<int32_t>(k);
GroupTopkParam.n = static_cast<uint16_t>(n);
GroupTopkParam.groupMultiFlag = static_cast<GroupTopkParam::GroupMultiFlag>(0);
if (n > 1) {
GroupTopkParam.groupMultiFlag = static_cast<GroupTopkParam::GroupMultiFlag>(1);
}
at::Tensor out = self;
auto idx = at::arange(1024, self.options().device(at::Device(at::kPrivateUse1)).dtype(at::kInt));
ParamSetter paramsetter;
paramsetter.Input(self, true)
.Input(idx, true)
.Output(out);
auto opGroupTopk = GroupTopkParamCache.getOperation(GroupTopkParam, "GroupTopkOperation");
RunAtbCmd(opGroupTopk, paramsetter, "GroupTopkOperation");
return;
}
namespace {
TORCH_LIBRARY_FRAGMENT(atb, m)
{
m.def("_npu_group_topk(Tensor self, int k=0, int group_num=1, int n=1) -> ()");
}
}
namespace {
TORCH_LIBRARY_IMPL(atb, PrivateUse1, m)
{
m.impl("_npu_group_topk", TORCH_FN(atb::_npu_group_topk));
}
}
}