bug: Poor GPU thread blocking for at-points operators with all `CEED_EVAL_NONE` inputs and outputs

[zatkins-dev]

We currently have poor GPU thread blocking for at-points operators with all CEED_EVAL_NONE inputs and outputs.

The issue is in backends/cuda-gen/ceed-cuda-gen-operator-build.cpp:CeedOperatorBuildKernel_Cuda_gen:

  // Lines 1228-1234
  if (Q_1d == 0) {
    if (is_at_points) Q_1d = max_num_points;
    else CeedCallBackend(CeedOperatorGetNumQuadraturePoints(op, &Q_1d));
  }
  if (Q == 0) Q = Q_1d;
  data->Q    = Q;
  data->Q_1d = Q_1d;

Rather than blocking by cell/slices of cells, we use a 1D blocking strategy with Q_1d = max_num_points. This can hit the maximum thread count/block size limits or the shared memory limits, causing a launch-time failure.

[zatkins-dev]

For max_num_points >= 512, blocking ends up looking like

• grid: num_cells
• block: [max_num_points 1 1]

In reality, we should probably use the basis for the point coordinates to define the blocking, since that will at least result in reasonable scaling

[jeremylt]

Yup, that should be the fix - use the blocking as if we actually cared about the basis for the operation