#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
using npu_utils = at_npu::native::NpuUtils;
namespace {
at::Tensor& nonzero_out_npu_nocheck(at::Tensor& result, const at::Tensor& self) {
c10::SmallVector<int64_t, N> output_sync_idx = {0};
at_npu::native::OpCommand cmd;
cmd.Sync(output_sync_idx)
.Name("NonZero")
.Input(self)
.Output(result)
.Attr("transpose", false)
.Run();
return result;
}
}
at::Tensor& nonzero_out(const at::Tensor& self, at::Tensor& result) {
auto output_size = op_infer::nonzero_npu_max_output_size(self);
if (self.numel() == 1 && self.dim() == 0) {
if (self.is_nonzero()) {
output_size = {1, 0};
} else {
output_size = {0, 0};
}
}
npu_preparation::CheckOut(
{self},
result,
npu_preparation::get_tensor_npu_format(self),
at::ScalarType::Long,
output_size);
if (!npu_utils::check_match(&result)) {
at::Tensor contiguous_result = npu_utils::format_contiguous(result);
nonzero_out_npu_nocheck(contiguous_result, self);
npu_utils::format_fresh_view(result, contiguous_result);
} else {
nonzero_out_npu_nocheck(result, self);
}
return result;
}
at::Tensor nonzero(const at::Tensor& self) {
auto output_size = op_infer::nonzero_npu_max_output_size(self);
if (self.numel() == 1 && self.dim() == 0) {
if (self.is_nonzero()) {
output_size = {1, 0};
} else {
output_size = {0, 0};
}
}
at::Tensor result = npu_preparation::apply_tensor(output_size, self.options().dtype(at::kLong), self);
nonzero_out_npu_nocheck(result, self);
return result;
}
}