#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
using calcu_op_util = at_npu::native::CalcuOpUtil;
using npu_utils = at_npu::native::NpuUtils;
namespace {
void nll_loss_forward_check(const at::Tensor& self, const at::Tensor& target) {
TORCH_CHECK(
self.dim() > 0 && self.dim() <= 2, "input tensor should be 1D or 2D"
+ OPS_ERROR(ErrCode::PARAM));
TORCH_CHECK(
target.dim() <= 1,
"0D or 1D target tensor expected, multi-target not supported"
+ OPS_ERROR(ErrCode::PARAM));
auto no_batch_dim = self.dim() == 1 && target.dim() == 0;
TORCH_CHECK(
no_batch_dim || self.size(0) == target.size(0),
"size mismatch (got input: ",
self.sizes(),
", target: ",
target.sizes(),
")"
+ OPS_ERROR(ErrCode::PARAM));
}
std::tuple<at::Tensor&, at::Tensor&> nll_loss_forward_out_nocheck(
at::Tensor& result,
at::Tensor& total_weight,
const at::Tensor& self,
const at::Tensor& target,
const at::Tensor& weight,
int64_t reduction,
int64_t ignore_index)
{
at::Tensor weight_tensor;
if (weight.defined()) {
weight_tensor = npu_utils::format_contiguous(weight);
} else {
weight_tensor = at::ones(self.size(1), self.options());
}
if (ignore_index >= 0 && ignore_index < self.size(-1)) {
at::Tensor zero = at::zeros(1, self.options());
calcu_op_util::AclrtMemcpyAsync({weight_tensor, ignore_index}, weight_tensor.itemsize(),
{zero, 0}, weight_tensor.itemsize(), ACL_MEMCPY_DEVICE_TO_DEVICE);
}
string reduction_str = op_plugin::utils::get_reduction_str(reduction);
auto scalar_type = target.scalar_type();
TORCH_CHECK((scalar_type == at::kLong || scalar_type == at::kInt),
"Expected object of scalar type ", at::kLong, " or ", at::kInt,
" but got scalar type ", scalar_type, " for argument 'target' in call to nll_loss_forward"
+ OPS_ERROR(ErrCode::TYPE));
at::Tensor target_cast = (scalar_type == at::kLong) ? at_npu::native::custom_ops::_npu_dtype_cast(target, at::kInt) : target;
at_npu::native::OpCommand cmd;
cmd.Name("NLLLoss")
.Input(self)
.Input(target_cast)
.Input(weight_tensor)
.Output(result)
.Output(total_weight)
.Attr("reduction", reduction_str)
.Attr("ignore_index", ignore_index)
.Run();
return std::tuple<at::Tensor&, at::Tensor&>(result, total_weight);
}
}
std::tuple<at::Tensor&, at::Tensor&> nll_loss_forward_out(
const at::Tensor& self,
const at::Tensor& target,
const c10::optional<at::Tensor>& weight_opt,
int64_t reduction,
int64_t ignore_index,
at::Tensor& result,
at::Tensor& total_weight) {
nll_loss_forward_check(self, target);
at::Tensor self_cp = self.dim() == 1 ? self.unsqueeze(0) : self;
const at::Tensor& weight = c10::value_or_else(weight_opt, [] {return at::Tensor();});
at::Tensor weight_tensor;
if (weight.defined()) {
weight_tensor = npu_utils::format_contiguous(weight);
} else {
auto options = self_cp.options();
weight_tensor = acl_op::ones(
self_cp.size(1),
c10::optTypeMetaToScalarType(options.dtype_opt()),
options.layout_opt(),
options.device_opt(),
options.pinned_memory_opt());
}
c10::SmallVector<int64_t, SIZE> output_size = {};
if (reduction == at::Reduction::None) {
output_size = {self_cp.size(0)};
}
npu_preparation::CheckOut(
{self_cp, target, weight_tensor},
result,
ACL_FORMAT_ND,
self_cp.scalar_type(),
output_size);
npu_preparation::CheckOut(
{self_cp, target, weight_tensor},
total_weight,
ACL_FORMAT_ND,
self_cp.scalar_type(),
{});
bool result_match = npu_utils::check_match(&result);
bool total_weight_match = npu_utils::check_match(&total_weight);
if (!(result_match && total_weight_match)) {
at::Tensor contiguous_result = result_match ? result : npu_utils::format_contiguous(result);
at::Tensor contiguous_total_weight =
total_weight_match ? total_weight : npu_utils::format_contiguous(total_weight);
nll_loss_forward_out_nocheck(contiguous_result, contiguous_total_weight, self_cp,
target, weight, reduction, ignore_index);
if (!result_match) {
npu_utils::format_fresh_view(result, contiguous_result);
}
if (!total_weight_match) {
npu_utils::format_fresh_view(total_weight, contiguous_total_weight);
}
} else {
nll_loss_forward_out_nocheck(result, total_weight, self_cp, target, weight, reduction, ignore_index);
}
if (self.dim() == 1 && reduction == at::Reduction::None) {
result.squeeze_(0);
}
return std::tie(result, total_weight);
}
std::tuple<at::Tensor, at::Tensor> nll_loss_forward(
const at::Tensor& self,
const at::Tensor& target,
const c10::optional<at::Tensor>& weight_opt,
int64_t reduction,
int64_t ignore_index) {
nll_loss_forward_check(self, target);
at::Tensor self_cp = self.dim() == 1 ? self.unsqueeze(0) : self;
c10::SmallVector<int64_t, SIZE> output_size = {};
c10::SmallVector<int64_t, SIZE> total_weight_size = {};
const at::Tensor& weight = c10::value_or_else(weight_opt, [] {return at::Tensor();});
if (reduction == at::Reduction::None) {
output_size = {self_cp.size(0)};
}
at::Tensor result = npu_preparation::apply_tensor_with_format(
self_cp, output_size, ACL_FORMAT_ND);
at::Tensor total_weight = npu_preparation::apply_tensor_with_format(
self_cp, total_weight_size, ACL_FORMAT_ND);
nll_loss_forward_out_nocheck(result, total_weight, self_cp,
target, weight, reduction, ignore_index);
if (self.dim() == 1 && reduction == at::Reduction::None) {
result.squeeze_(0);
}
return std::tuple<at::Tensor, at::Tensor>(result, total_weight);
}
}