#include "op_plugin/AclOpsInterface.h"
#include "op_plugin/utils/OpAdapter.h"
namespace acl_op {
using npu_preparation = at_npu::native::OpPreparation;
std::tuple<at::Tensor, at::Tensor> batch_norm_stats(const at::Tensor& self, double eps)
{
TORCH_CHECK(
self.ndimension() >= 2,
"Expected 2D+ Tensor, but got tensor with ",
self.ndimension(),
" Dimension" + OPS_ERROR(ErrCode::PARAM));
auto output_size = {self.size(1)};
at::Tensor mean = npu_preparation::apply_tensor(output_size, self.options().dtype(at::kFloat), self);
at::Tensor invstd = npu_preparation::apply_tensor(output_size, self.options().dtype(at::kFloat), self);
c10::SmallVector<int64_t, N> dim;
int dimN = self.ndimension();
for (int i = 0; i < dimN; i++) {
if (i == 1) {
continue;
}
dim.emplace_back(i);
}
at::Tensor self_copy = self;
if (self.scalar_type() != at::kFloat) {
self_copy = at_npu::native::custom_ops::_npu_dtype_cast(self_copy, at::kFloat);
}
at_npu::native::OpCommand cmd_mean;
cmd_mean.Name("ReduceMean")
.Input(self_copy)
.Input(dim, at::kInt)
.Output(mean)
.Attr("keep_dims", (bool) false)
.Run();
at::Tensor mean_copy = mean;
if (mean.dim() != 0) {
auto dim_vector = op_infer::array_to_small_vector(dim);
for (uint64_t i = 0; i < dim_vector.size(); i++) {
mean_copy = mean_copy.unsqueeze(dim_vector[i]);
}
}
mean_copy = mean_copy.expand(self.sizes());
at_npu::native::OpCommand cmd_invstd;
cmd_invstd.Name("ReduceStdWithMean")
.Input(self_copy)
.Input(mean_copy)
.Output(invstd)
.Attr("dim", dim)
.Attr("unbiased", false)
.Attr("keepdim", false)
.Attr("invert", true)
.Attr("epsilon", static_cast<float>(eps))
.Run();
return std::tie(mean, invstd);
}
}