Function HasOptimizedOneDNNKernel

paddle/phi/kernels/reduce_kernel_impl.cc:23–55

oneDNN's reduction kernel is optimized only for reducing throughout the most outer dims, so in case of another type of reduction, it would be better to fallback to native implementation

// most outer dims, so in case of another type of reduction, it would be
// better to fallback to native implementation
inline bool HasOptimizedOneDNNKernel(const KernelContext* dev_ctx) {
  const DenseTensor& x = dev_ctx->InputAt<DenseTensor>(0);
  IntArray dims_array;
  const TensorRef& dims_tmp = dev_ctx->AttrAt<TensorRef>(0);
  dims_array = IntArray(*dims_tmp.Get());
  int ndims = x.dims().size();
  const bool reduce_all = recompute_reduce_all(x, dims_array);
  auto dims = dims_array.GetData();

  // native reduce kernels don't support bf16
  // so oneDNN kernel is enforced in that case
  if (x.dtype() == phi::DataType::BFLOAT16) return true;

  if (reduce_all) {
    return true;
  }

  for (auto& dim : dims) {
    if (dim < 0) {
      dim += ndims;
    }
  }

  sort(dims.begin(), dims.end());

  for (size_t i = 0; i < dims.size(); ++i) {
    if (dims[dims.size() - i - 1] != static_cast<int>(ndims - i - 1)) {
      return false;
    }
  }

  return true;
}

bool ReduceCheckIfOneDNNSupport(const KernelContext* dev_ctx) {
  if (dev_ctx->InputAt<DenseTensor>(0).dims().size() > 5 ||