No valid engine configs for NCHW sigmoid

[swimvtec]

Describe the bug
No valid engine configuration is found for a simple sigmoid activation with NCHW layout. Works with the legacy API.

Expected behavior
Simple graphs just work. No need to convert tensor shapes.

System Environment (please complete the following information):

• cudnn_frontend version: 1.12
• cudnn_backend version: 9.11
• GPU arch: RTX A2000
• cuda runtime version: 12.8
• cuda driver version: 570.172.08
• host compiler: clang19
• OS: ubuntu22.04

API logs
API logs are attached:

be.log
fe.log

To Reproduce

static void SetTensorShape(
    std::shared_ptr<cudnn_frontend::graph::Tensor_attributes>& tensor,
    Shape const&                                               shape)
{
  tensor->set_dim({shape.batch_size, shape.depth, shape.height, shape.width})
      .set_stride({shape.height * shape.width * shape.depth,
                   shape.height * shape.width, shape.width, 1});
}

static std::shared_ptr<cudnn_frontend::graph::Tensor_attributes>
TensorFromShape(cudnn_frontend::graph::Graph* graph, Shape const& shape,
                std::string const& name)
{
  namespace fe = cudnn_frontend;

  std::shared_ptr<fe::graph::Tensor_attributes> tensor =
      graph->tensor(fe::graph::Tensor_attributes().set_name(name));

  SetTensorShape(tensor, shape);

  return tensor;
}

TEST(Sigmoid, Small)
{
  ... code to get cudnn handle ...

  namespace fe                = cudnn_frontend;
  auto              fwd_graph = std::make_unique<fe::graph::Graph>();
  fe::graph::Graph* graph     = fwd_graph.get();

  graph->set_io_data_type(fe::DataType_t::FLOAT)
      .set_compute_data_type(fe::DataType_t::FLOAT)
      .set_intermediate_data_type(fe::DataType_t::FLOAT);

  Shape input_shape{2, 1, 1, 2};

  auto tensor_x = TensorFromShape(graph, input_shape, "input");

    auto pointwise_attribs =
      fe::graph::Pointwise_attributes().set_compute_data_type(
          fe::DataType_t::FLOAT).set_mode(fe::PointwiseMode_t::SIGMOID_FWD);
  auto tensor_y = graph->pointwise(tensor_x, pointwise_attribs);

  SetTensorShape(tensor_y, input_shape);
  tensor_y->set_output(true);

  if (auto err = graph->validate(); err.is_bad())
  {
    GTEST_FAIL();
  }
  if (auto err = graph->build_operation_graph(cudnn_handle); err.is_bad())
  {
    GTEST_FAIL();
  }

  if (auto err = graph->create_execution_plans(
          {fe::HeurMode_t::B, fe::HeurMode_t::FALLBACK});
      err.is_bad())
  {
    GTEST_FAIL();
  }
}

Additional context
Does not work with NHWC layout as well. Seems to work reliably with 3 dimensional tensors only.

[swalters22]

Modern GPU architectures are expected to support NHWC more performantly than NCHW, so that is the development focus. It is correct that NCHW sigmoid activation was supported in the legacy API, but as the graph API is more forward looking, it is focused on NHWC workloads. Here, unfortunately, it is expected that NCHW is not supported in the graph API.

This code seems to be additionally victimized by an undocumented ( or non-obvious ) constraint that channels must be > 1 to support simple NHWC pointwise operations.

Are you able to share the larger graphs or workloads whose support you are looking for with these and other pointwise operations? It maybe that with some idea of the larger graph, cuDNN can offer better performance with fused workloads for those graphs.