ROperator_Pool doesn't override any of the Generate_GPU_* methods, so the alpaka backend falls through to the base class and silently emits empty kernels for MaxPool nodes. The CPU side has the full implementation though (1D, 2D, 3D, max and average pool).
GPU codegen should be added for MaxPool 1D,2D and 3D. A one thread per-output kernel with constexpr baked attributes gives a working baseline that further GPU side optimizations can build on to
ROperator_Pooldoesn't override any of the Generate_GPU_* methods, so the alpaka backend falls through to the base class and silently emits empty kernels for MaxPool nodes. The CPU side has the full implementation though (1D, 2D, 3D, max and average pool).GPU codegen should be added for MaxPool 1D,2D and 3D. A one thread per-output kernel with constexpr baked attributes gives a working baseline that further GPU side optimizations can build on to