concatwrite

func Apply ¶

func Apply(g *graph.Graph) int

Apply rewrites the graph in place, collapsing every (producer-Conv → Concat) tail pair where the producer's output is the LAST input of the Concat AND has exactly one consumer (the Concat itself) into a single fused "ConcatFusedTail" node. The fused node allocates the joint output buffer once, copies all prefix inputs into the head of that buffer with the same block-copy ordering Concat uses, and dispatches the producer Conv kernel to write its output directly into the tail slice — eliminating the producer's standalone output materialisation AND the corresponding memcpy that an unfused Concat would perform on that tail slice.

Why "tail-only" rather than full N-producer fusion? In Real-ESRGAN's RDB dense-connection pattern (276 Concat ops out of 1094 total nodes post other-fusion), each Concat reads K producers of which only the LAST is freshly produced and single-consumer; the K-1 earlier producers are tensors reused across multiple downstream Concats. They cannot be folded into the fused buffer (folding would require a separate dedicated buffer per consumer), but the latest one — the dense-block's just-emitted feature map — is always single-use and is the largest-payload candidate for write fusion. Empirically the pass folds ~276 Concats on Real-ESRGAN x4 (one per Concat in the dense blocks), 0 on realesr-general-x4v3 (no Concats), and 0 on SPAN.

Eligibility constraints (verified for every candidate before fusing):

1. The Concat node has axis == 1 (the channel axis in NCHW).
2. The Concat node has at least 2 inputs.
3. The Concat's LAST input is produced by a node N whose OpType is one of "Conv", "ConvLeakyRelu", "ConvPRelu" — the operators whose Execute methods take their output buffer from the supplied allocator.
4. Producer N's output is consumed exactly once in the entire graph.
5. Producer N appears immediately before the Concat in topological order.

Apply returns the number of Concat nodes that were rewritten.