#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "op_plugin/utils/KernelNpuOutputSize.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
at::Tensor& var_out(
const at::Tensor& self,
at::OptionalIntArrayRef dim,
const c10::optional<c10::Scalar>& correction,
bool keepdim,
at::Tensor& out)
{
if (!correction_fits_aclnn_int64(correction)) {
at::Tensor cpu_out = out.cpu();
at::var_out(cpu_out, self.cpu(), dim, correction, keepdim);
out.copy_(cpu_out);
return out;
}
DO_COMPATIBILITY(aclnnVarCorrection, acl_op::var_out(self, dim, correction, keepdim, out));
c10::SmallVector<int64_t, op_infer::SIZE> real_dim = {};
if (dim.has_value()) {
real_dim = op_infer::array_to_small_vector(dim.value());
}
auto output_size = op_infer::reduce_ops_npu_output_size(self, real_dim, keepdim);
int64_t real_correction = correction.has_value() ? correction.value().toLong() : 1;
at_npu::native::OpPreparation::check_tensor(
{self},
out,
self,
output_size);
auto rd = at::IntArrayRef(real_dim);
EXEC_NPU_CMD(aclnnVarCorrection, self, rd, real_correction, keepdim, out);
return out;
}
at::Tensor var(
const at::Tensor & self,
at::OptionalIntArrayRef dim,
const c10::optional<c10::Scalar>& correction,
bool keepdim)
{
if (!correction_fits_aclnn_int64(correction)) {
return at::var(self.cpu(), dim, correction, keepdim).to(self.options());
}
DO_COMPATIBILITY(aclnnVarCorrection, acl_op::var(self, dim, correction, keepdim));
c10::SmallVector<int64_t, op_infer::SIZE> real_dim = {};
if (dim.has_value()) {
real_dim = op_infer::array_to_small_vector(dim.value());
}
auto output_size = op_infer::reduce_ops_npu_output_size(self, real_dim, keepdim);
int64_t real_correction = correction.has_value() ? correction.value().toLong() : 1;
auto result = npu_preparation::apply_tensor_without_format(output_size, self.options());
at_npu::native::OpPreparation::check_tensor(
{self},
result,
self,
output_size);
auto rd = at::IntArrayRef(real_dim);
EXEC_NPU_CMD(aclnnVarCorrection, self, rd, real_correction, keepdim, result);
return result;
}
std::tuple<at::Tensor, at::Tensor> var_mean(
const at::Tensor& self,
at::OptionalIntArrayRef dim,
const c10::optional<c10::Scalar>& correction,
bool keepdim)
{
if (!correction_fits_aclnn_int64(correction)) {
auto cpu_tup = at::var_mean(self.cpu(), dim, correction, keepdim);
return std::make_tuple(
std::get<0>(cpu_tup).to(self.options()),
std::get<1>(cpu_tup).to(self.options()));
}
c10::SmallVector<int64_t, N> real_dim = op_plugin::utils::get_dimlist_for_tensor(self);
if (dim.has_value()) {
real_dim = op_infer::array_to_small_vector(dim.value());
}
int64_t real_correction = correction.has_value() ? correction.value().toLong() : 1;
auto output_size = op_infer::reduce_ops_npu_output_size(self, real_dim, keepdim);
auto var = npu_preparation::apply_tensor_without_format(output_size, self.options());
auto mean = npu_preparation::apply_tensor_without_format(output_size, self.options());
auto rd = at::IntArrayRef(real_dim);
EXEC_NPU_CMD(aclnnVarMean, self, rd, real_correction, keepdim, var, mean);
return std::tuple<at::Tensor, at::Tensor>(var, mean);
}
}