Deeploy.CommonExtensions.OptimizationPasses.TopologyOptimizationPasses.LoweringOptimizationPasses.isPULPStemConv

Deeploy.CommonExtensions.OptimizationPasses.TopologyOptimizationPasses.LoweringOptimizationPasses.isPULPStemConv(node: Node) → bool

First-layer convolution that can stay channels-first on both sides.

Both ends have to want that layout: the input must be a graph input, which arrives channels-first, and the consumer must be a depthwise convolution, which reads channels-first. Feeding a dense convolution instead would only move the transpose rather than remove it. The shape is the one PULPStemConv3x3.c has a fast path for; anything else would land in its reference loop, which is far slower than the im2col kernel it replaced.