#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;
inline void alpha_check_npu(const at::ScalarType dtype, at::Scalar alpha)
{
TORCH_CHECK(!alpha.isBoolean() || dtype == at::ScalarType::Bool,
"Boolean alpha only supported for Boolean results." + OPS_ERROR(ErrCode::TYPE));
TORCH_CHECK(isFloatingType(dtype) || isComplexType(dtype) || alpha.isIntegral(true),
"For integral input tensors, argument alpha must not be a floating point number."
+ OPS_ERROR(ErrCode::TYPE));
}
static at::Tensor self_tensor_to_device(const at::Tensor &tensor, const at::ScalarType result_type,
const c10::Device device)
{
if (npu_preparation::is_scalar_wrapped_to_tensor(tensor) ||
(tensor.dim() == 0 && !torch_npu::utils::is_npu(tensor))) {
at::Scalar scalar = tensor.item();
return npu_preparation::copy_scalar_to_device(scalar, result_type, device);
}
return tensor;
}
static at::Tensor &add_out_npu_nocheck(
const at::Tensor &self,
const at::Tensor &other,
const at::Scalar &alpha,
at::Tensor &result)
{
if (other.dim() == 0 && !torch_npu::utils::is_npu(other)) {
c10::Scalar others = other.item();
EXEC_NPU_CMD(aclnnAdds, self, others, alpha, result);
} else {
if (self.dim() == 0 && !torch_npu::utils::is_npu(self)) {
static const bool is_aclnn_available = check_aclnn_kernel_available("aclnnAddV3");
if (is_aclnn_available && self.dtype() != at::kBool) {
c10::Scalar selfs = self.item();
EXEC_NPU_CMD(aclnnAddV3, selfs, other, alpha, result);
} else {
at::Tensor self_cp = self_tensor_to_device(self, result.scalar_type(), result.device());
EXEC_NPU_CMD(aclnnAdd, self_cp, other, alpha, result);
}
} else {
EXEC_NPU_CMD(aclnnAdd, self, other, alpha, result);
}
}
return result;
}
static at::Tensor &inplace_add_out_npu_no_check(at::Tensor &self, const at::Tensor &other, const at::Scalar &alpha)
{
if (other.dim() == 0 && !torch_npu::utils::is_npu(other)) {
c10::Scalar other_scalar = other.item();
EXEC_NPU_CMD(aclnnInplaceAdds, self, other_scalar, alpha);
} else {
EXEC_NPU_CMD(aclnnInplaceAdd, self, other, alpha);
}
return self;
}
static at::Tensor add_dest_output(const at::Tensor &self, const at::Tensor &other)
{
bool isSelfWrapped = npu_preparation::is_scalar_wrapped_to_tensor(self);
return isSelfWrapped ? other : self;
}
at::Tensor add(const at::Tensor &self, const at::Tensor &other, const at::Scalar &alpha)
{
DO_COMPATIBILITY(aclnnAdd, acl_op::add(self, other, alpha));
DO_COMPATIBILITY(aclnnAdds, acl_op::add(self, other, alpha));
std::vector<at::Tensor> tensor_list = {self, other};
auto maybe_names = op_plugin::utils::compute_names_npu(tensor_list);
at::Tensor output_tensor = add_dest_output(self, other);
auto output_size = op_infer::broadcast_ops_npu_output_size(self, other);
at::ScalarType result_type = at::native::result_type(self, other);
alpha_check_npu(result_type, alpha);
at::Tensor result =
npu_preparation::apply_tensor_without_format(output_size, output_tensor.options().dtype(result_type));
add_out_npu_nocheck(self, other, alpha, result);
at::namedinference::propagate_names_if_nonempty(result, maybe_names);
return result;
}
at::Tensor add(const at::Tensor &self, const at::Scalar &other, const at::Scalar &alpha)
{
DO_COMPATIBILITY(aclnnAdds, acl_op::add(self, other, alpha));
auto output_size = op_infer::input_same_output_size(self);
at::ScalarType result_type = at::native::result_type(self, other);
alpha_check_npu(result_type, alpha);
at::Tensor result = npu_preparation::apply_tensor_without_format(output_size, self.options().dtype(result_type));
EXEC_NPU_CMD(aclnnAdds, self, other, alpha, result);
at::namedinference::propagate_names(result, self);
return result;
}
at::Tensor &add_out(const at::Tensor &self, const at::Tensor &other, const at::Scalar &alpha, at::Tensor &result)
{
DO_COMPATIBILITY(aclnnAdd, acl_op::add_out(self, other, alpha, result));
DO_COMPATIBILITY(aclnnAdds, acl_op::add_out(self, other, alpha, result));
std::vector<at::Tensor> tensor_list = {self, other};
auto maybe_names = op_plugin::utils::compute_names_npu(tensor_list);
bool isSelfWrapped = npu_preparation::is_scalar_wrapped_to_tensor(self);
at::Tensor output_tensor = isSelfWrapped ? other : self;
auto output_size = op_infer::broadcast_ops_npu_output_size(self, other);
at::ScalarType result_type = at::native::result_type(self, other);
npu_preparation::check_tensor({self}, result, result, output_size);
npu_preparation::check_memory({self, other}, {result});
add_out_npu_nocheck(self, other, alpha, result);
at::namedinference::propagate_names_if_nonempty(result, maybe_names);
return result;
}
at::Tensor &add_(at::Tensor &self, const at::Tensor &other, const at::Scalar &alpha)
{
DO_COMPATIBILITY(aclnnInplaceAdd, acl_op::add_(self, other, alpha));
DO_COMPATIBILITY(aclnnInplaceAdds, acl_op::add_(self, other, alpha));
std::vector<at::Tensor> tensor_list = {self, other};
auto maybe_names = op_plugin::utils::compute_names_npu(tensor_list);
npu_preparation::check_memory({self, other}, {self});
inplace_add_out_npu_no_check(self, other, alpha);
at::namedinference::propagate_names_if_nonempty(self, maybe_names);
return self;
}
at::Tensor &add_(at::Tensor &self, const at::Scalar &other, const at::Scalar &alpha)
{
DO_COMPATIBILITY(aclnnInplaceAdds, acl_op::add_(self, other, alpha));
EXEC_NPU_CMD(aclnnInplaceAdds, self, other, alpha);
return self;
}
}