#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;
namespace{
float calculate_p(c10::optional<at::Scalar> p)
{
if (p.has_value()) {
float val = op_plugin::utils::get_scalar_float_value(p.value());
if (val == INFINITY) {
return static_cast<float>(INT_MAX);
} else if (val == -INFINITY) {
return static_cast<float>(INT_MIN);
} else {
return p.value().toFloat();
}
} else {
return static_cast<float>(2.0);
}
}
inline bool check_use_aclop(float pfloat)
{
if (c10_npu::IsAclnnOnly() || op_plugin::utils::is_gte_cann_version_900()) {
return false;
}
if (pfloat != 0.0 && pfloat != 1.0 && pfloat != 2.0 && pfloat != 3.0) {
if (op_plugin::utils::is_gte_cann_version_810rc1() &&
(pfloat == static_cast<float>(INT_MAX) || pfloat == static_cast<float>(INT_MIN))) {
return false;
}
return true;
}
return false;
}
inline at::Tensor &norm_out_npu_nocheck_opapi(at::Tensor &out,
const at::Tensor &self,
c10::optional<at::Scalar> p,
at::IntArrayRef dim,
bool keepdim)
{
at::Scalar pvalue = 2;
if (p.has_value()) {
pvalue = p.value();
}
EXEC_NPU_CMD(aclnnNorm, self, pvalue, dim, keepdim, out);
return out;
}
inline at::Tensor &norm_out_imp(const at::Tensor &self,
const c10::optional<at::Scalar> &p,
at::IntArrayRef dim,
bool keepdim,
at::ScalarType dtype,
at::Tensor &out)
{
float pfloat = calculate_p(p);
if (check_use_aclop(pfloat)) {
return acl_op::norm_out(self, p, dim, keepdim, dtype, out);
} else {
auto outputSize = op_infer::reduce_ops_npu_output_size(self, dim, keepdim);
npu_preparation::check_tensor({self}, out, dtype, outputSize);
return norm_out_npu_nocheck_opapi(out, self, p, dim, keepdim);
}
}
inline at::Tensor norm_imp(const at::Tensor &self,
const c10::optional<at::Scalar> &p,
at::IntArrayRef dim,
bool keepdim,
at::ScalarType dtype)
{
float pfloat = calculate_p(p);
if (check_use_aclop(pfloat)) {
return acl_op::norm(self, p, dim, keepdim, dtype);
} else {
auto outputSize = op_infer::reduce_ops_npu_output_size(self, dim, keepdim);
auto dtype_checked = dtype;
if (self.is_complex()) {
dtype_checked = self.scalar_type() == at::kComplexFloat ? at::kFloat : at::kDouble;
}
at::Tensor out = npu_preparation::apply_tensor_with_sizes(outputSize, self.options().dtype(dtype_checked));
return norm_out_npu_nocheck_opapi(out, self, p, dim, keepdim);
}
}
}
at::Tensor& norm_out(const at::Tensor &self,
const c10::optional<at::Scalar> &p,
at::IntArrayRef dim,
bool keepdim,
at::ScalarType dtype,
at::Tensor &out)
{
DO_COMPATIBILITY(aclnnNorm, acl_op::norm_out(self, p, dim, keepdim, dtype, out));
return norm_out_imp(self, p, dim, keepdim, out.scalar_type(), out);
}
at::Tensor& norm_out(const at::Tensor &self,
const c10::optional<at::Scalar> &p,
at::IntArrayRef dim,
bool keepdim,
at::Tensor &out)
{
DO_COMPATIBILITY(aclnnNorm, acl_op::norm_out(self, p, dim, keepdim, out));
return norm_out_imp(self, p, dim, keepdim, out.scalar_type(), out);
}
at::Tensor norm(const at::Tensor &self,
const c10::optional<at::Scalar> &p,
at::IntArrayRef dim,
bool keepdim,
at::ScalarType dtype)
{
DO_COMPATIBILITY(aclnnNorm, acl_op::norm(self, p, dim, keepdim, dtype));
return norm_imp(self, p, dim, keepdim, dtype);
}
at::Tensor norm(const at::Tensor &self,
const c10::optional<at::Scalar> &p,
at::ScalarType dtype)
{
DO_COMPATIBILITY(aclnnNorm, acl_op::norm(self, p, dtype));
return norm_imp(self, p, {}, false, dtype);
}
at::Tensor norm(const at::Tensor &self,
const at::Scalar &p)
{
DO_COMPATIBILITY(aclnnNorm, acl_op::norm(self, p));
return norm_imp(self, p, {}, false, self.scalar_type());
}
at::Tensor norm(const at::Tensor &self,
const c10::optional<at::Scalar> &p,
at::IntArrayRef dim,
bool keepdim)
{
DO_COMPATIBILITY(aclnnNorm, acl_op::norm(self, p, dim, keepdim));
return norm_imp(self, p, dim, keepdim, self.scalar_type());
}
}