Hailo10H Compiler: Layer Support for Input Tensors with Large Dimension Sizes

Hi @user976,

I noticed both threads (Hailo10H Data Flow Compiler Error: Spatial Reshape Infeasible) are related to the same EfficientLoFTR compilation, so I’ll address them together.

Both errors are compiler allocation failures where the intermediate tensor sizes (the 4800×4800 similarity matrix from matmul33, and the [1, 6400, 12, 12] reshape) exceed what the hardware can map for those specific operations at that resolution.

Since you confirmed that compilation succeeds at lower resolutions, you might try reducing the input to an intermediate size (e.g. 416×416 or 512×512) and see where the threshold is.

Alternatively, you could set end_node_names during parsing to cut the graph before the problematic layers and handle those operations in CPU post-processing.

Thanks,