#include <ATen/native/ForeachUtils.h>
#include "op_plugin/OpApiInterface.h"
#include "op_plugin/utils/op_api_common.h"
#include "op_plugin/utils/custom_functions/opapi/scalar_op_api.h"
#include "torch_npu/csrc/framework/utils/UtilForOpAdapter.h"
#include "op_plugin/AclOpsInterface.h"
namespace op_api {
using npu_preparation = at_npu::native::OpPreparation;
constexpr int FOREACH_NORM_MAX_TENSOR_COUNT = 24;
constexpr int L1_NORM = 1;
constexpr int L2_NORM = 2;
bool scalarProcess(const at::Scalar& scalar, double *p)
{
if (scalar.isIntegral(false)) {
*p = scalar.to<int64_t>();
} else if (scalar.isFloatingPoint()) {
*p = scalar.to<double>();
} else {
TORCH_CHECK(false, "foreach_tensor_norm_npu expects scalar to be integer or float", OPS_ERROR(ErrCode::TYPE));
}
return true;
}
bool checkData(const at::TensorList self, const at::Scalar& scalar)
{
double p = 0.0;
scalarProcess(scalar, &p);
const bool has_int_or_complex = std::any_of(self.begin(), self.end(), [](const auto &t) {
const auto scalar_type = t.scalar_type();
return at::isIntegralType(scalar_type, true) || at::isComplexType(scalar_type);
});
return !at::native::can_use_fast_route(self) || has_int_or_complex ||
!(p == static_cast<double>(L1_NORM) || p == static_cast<double>(L2_NORM));
}
bool checkDataWithInfinit(const at::TensorList self, const at::Scalar& scalar)
{
double p = 0.0;
static bool isRegBaseSoc = c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend950;
scalarProcess(scalar, &p);
const bool has_int_or_complex = std::any_of(self.begin(), self.end(), [](const auto &t) {
const auto scalar_type = t.scalar_type();
return at::isIntegralType(scalar_type, true) || at::isComplexType(scalar_type);
});
if (isRegBaseSoc) {
return !at::native::can_use_fast_route(self) || has_int_or_complex ||
!(p == static_cast<double>(L1_NORM) || p == static_cast<double>(L2_NORM) ||
p == std::numeric_limits<double>::infinity());
} else {
return !at::native::can_use_fast_route(self) || has_int_or_complex ||
!(p == static_cast<double>(L1_NORM) || p == static_cast<double>(L2_NORM));
}
}
void _split_and_exec_npu_cmd_norm(at::TensorList tensors, const at::Scalar& scalar, at::TensorList result_list)
{
size_t tensor_count = tensors.size();
size_t max_tensor_count = FOREACH_NORM_MAX_TENSOR_COUNT;
size_t loop_time = tensor_count / max_tensor_count;
if (tensor_count <= max_tensor_count) {
EXEC_NPU_CMD(aclnnForeachNorm, tensors, scalar, result_list);
return;
}
for (size_t i = 0; i < loop_time; i++) {
at::TensorList temp_tensors1(tensors.data() + i * max_tensor_count, max_tensor_count);
at::TensorList temp_result(result_list.data() + i * max_tensor_count, max_tensor_count);
EXEC_NPU_CMD(aclnnForeachNorm, temp_tensors1, scalar, temp_result);
}
size_t remaining_count = tensor_count % max_tensor_count;
if (remaining_count) {
at::TensorList temp_tensors1(tensors.data() + loop_time * max_tensor_count, remaining_count);
at::TensorList temp_result(result_list.data() + loop_time * max_tensor_count, remaining_count);
EXEC_NPU_CMD(aclnnForeachNorm, temp_tensors1, scalar, temp_result);
}
}
#if VERSION_BETWEEN(V2R1, V2R3)
std::vector<at::Tensor> _foreach_norm(const at::TensorList self, const at::Scalar& scalar)
{
at::native::check_foreach_api_restrictions(self);
static const bool is_support_nd_out = (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1 &&
c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend310B1 ||
c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend950);
if (!is_support_nd_out) {
return at::native::foreach_tensor_norm_slow(self, scalar);
}
DO_COMPATIBILITY(aclnnForeachNorm, at::native::foreach_tensor_norm_slow(self, scalar));
if (checkData(self, scalar)) {
return at::native::foreach_tensor_norm_slow(self, scalar);
}
auto scalar_type = self[0].scalar_type();
std::vector<at::Tensor> result;
for (const at::Tensor &tensor : self) {
result.push_back(npu_preparation::apply_tensor_without_format({}, tensor.options().dtype(scalar_type)));
}
at::TensorList result_ = at::TensorList(result);
_split_and_exec_npu_cmd_norm(self, scalar, result_);
return result;
}
#endif
#if VERSION_BETWEEN(V2R4, VERSION_NEWEST)
std::vector<at::Tensor> _foreach_norm(const at::TensorList self, const at::Scalar& scalar, at::optional<at::ScalarType> opt_dtype)
{
at::native::check_foreach_api_restrictions(self);
static const bool is_support_nd_out = (c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend910B1 &&
c10_npu::GetSocVersion() < c10_npu::SocVersion::Ascend310B1 ||
c10_npu::GetSocVersion() >= c10_npu::SocVersion::Ascend950);
if (!is_support_nd_out) {
return at::native::foreach_tensor_norm_slow(self, scalar, opt_dtype);
}
DO_COMPATIBILITY(aclnnForeachNorm, at::native::foreach_tensor_norm_slow(self, scalar, opt_dtype));
if (checkDataWithInfinit(self, scalar)) {
return at::native::foreach_tensor_norm_slow(self, scalar, opt_dtype);
}
auto scalar_type = self[0].scalar_type();
auto dtype = opt_dtype.has_value() ? opt_dtype.value() : scalar_type;
TORCH_CHECK(promoteTypes(scalar_type, dtype) == dtype, "foreach_norm", ": the dtype of the input ", "(",
scalar_type, ") should be convertible ", "without narrowing to the specified dtype (", dtype, ")", OPS_ERROR(ErrCode::TYPE));
std::vector<at::Tensor> inputs;
if (scalar_type != dtype) {
for (const at::Tensor& tensor :self) {
inputs.push_back(op_api::_npu_dtype_cast(tensor, dtype));
}
}
at::TensorList inputs_ = at::TensorList(inputs);
std::vector<at::Tensor> result;
for (const at::Tensor &tensor : self) {
result.push_back(npu_preparation::apply_tensor_without_format({}, tensor.options().dtype(dtype)));
}
at::TensorList result_ = at::TensorList(result);
if (scalar_type != dtype) {
_split_and_exec_npu_cmd_norm(inputs_, scalar, result_);
} else {
_split_and_exec_npu_cmd_norm(self, scalar, result_);
}
return result;
}
#endif
}