Fast and Fusiest Mapper (FFM) in DNN Accelerators

• FFM is a designation for DNNFuser’s rapid mapping approach that uses a Transformer to perform one-shot layer fusion optimization.
• It reformulates the combinatorial inter-layer mapping challenge as an autoregressive sequence generation, drastically reducing inference time.
• FFM minimizes data movement and energy consumption in DNN accelerators by optimizing fusion decisions with a single forward pass.

Searching arXiv for the cited DNNFuser paper and closely related context. arxiv.search({"query":"(Kao et al., 2022)", "max_results": 5, "sort_by": "submittedDate"}) arxiv.search({"query":"DNNFuser layer fusion mapper Transformer DNN accelerators", "max_results": 10, "sort_by": "relevance"}) arxiv.search({"query":"\"Fast and Fusiest Mapper\" OR FFM DNNFuser", "max_results": 10, "sort_by": "relevance"}) arxiv.search({"query":"layer fusion mapping DNN accelerator mapper search-based arXiv", "max_results": 10, "sort_by": "relevance"}) In DNN accelerator research, Fast and Fusiest Mapper (FFM) is best understood as a shorthand characterization of the capability realized by DNNFuser, rather than as a separate algorithmic family: a mapper that is both fast at inference time and highly effective at layer fusion in inter-layer map-space. DNNFuser frames layer-fusion mapping as autoregressive sequence generation with a generative pre-trained Transformer, replacing iterative exploration by a single forward pass that predicts a valid fusion/mapping decision sequence. The method is positioned as a one-shot inference-based alternative to search-based mappers and is reported to infer solutions with compatible performance while being 66x–127x faster (Kao et al., 2022).

1. Terminological scope and disambiguation

The acronym FFM is used with unrelated meanings across the literature. In the present context, it refers to the “Fast and Fusiest Mapper” characterization associated with DNNFuser’s objective of fast, high-quality layer-fusion mapping (Kao et al., 2022).

Source	Expansion of FFM	Meaning
(Kao et al., 2022)	Fast and Fusiest Mapper	Descriptive label for a fast, fusion-effective mapper realized by DNNFuser
(Luo et al., 2024)	Feature Fusion Module	Local feature fusion module for occluded pedestrian detection
(Okwata et al., 23 May 2026)	Five-Factor Model	Personality framework used for cybersecurity-awareness content routing

This distinction is substantive rather than terminological. In the pedestrian-detection work, FFM is a post-detection fusion/reconstruction stage that combines local body-part detections into whole-pedestrian boxes (Luo et al., 2024). In the cybersecurity-awareness study, FFM refers to the Five-Factor Model of personality, also called the Big Five or OCEAN, and has no relation to DNN accelerator mapping (Okwata et al., 23 May 2026). The DNNFuser usage is therefore specific to layer-fusion mapping and should not be conflated with either computer-vision feature fusion or personality modeling.

2. Inter-layer map-space and the problem FFM addresses

The problem addressed by FFM, as instantiated by DNNFuser, is inter-layer mapping, also described as layer-fusion map-space. The central motivation is that data movement dominates both latency and energy in DNN inference accelerators, so the decision of whether and how to fuse adjacent layers has direct consequences for compute and energy efficiency (Kao et al., 2022).

Layer fusion is valuable because it can reduce intermediate memory traffic, increase data reuse, lower off-chip bandwidth demand, reduce energy consumption, and often improve throughput. The difficulty is that the mapping problem is not limited to deciding whether two layers should be fused. A valid fusion plan must respect tensor shape compatibility, producer-consumer dependencies, accelerator resource constraints, and the legality of reusing on-chip buffers across fused layers. Each fusion decision constrains later scheduling and memory decisions, so the map-space is explicitly treated as combinatorial (Kao et al., 2022).

This framing explains why the “fast” aspect of FFM matters. Classical mapping approaches typically rely on expensive search over this combinatorial space. A plausible implication is that the practical bottleneck is not only mapping quality but also the runtime overhead of producing the map itself. DNNFuser’s contribution is to move that cost from inference-time exploration to training-time learning.

3. Representation of mapping as sequence generation

DNNFuser formulates layer-fusion optimization as a sequence modeling problem. The input encodes the DNN together with the accelerator/fusion context, including layer attributes and mapping-relevant features, and the output is a sequence of mapping decisions describing how layers should be fused and mapped (Kao et al., 2022).

The core probabilistic formulation is:

$$p(y \mid x) = \prod_{t=1}^{T} p(y_t \mid y_{<t}, x)$$

where $x$ is the encoded problem instance and $y = (y_1,\dots,y_T)$ is the output mapping/fusion sequence. In this formulation, each $y_t$ is a token representing a fusion or mapping decision. The learned mapper is therefore not a handcrafted rule engine or an explicit graph-search procedure; it is an autoregressive generator over mapping tokens (Kao et al., 2022).

The training objective is the standard negative log-likelihood over token sequences:

$$\mathcal{L} = - \sum_{t=1}^{T} \log p(y_t \mid y_{<t}, x)$$

This objective makes the Transformer learn expert or high-quality mapping sequences token by token. The representational choice is important: mapping decisions are encoded as tokens representing layer identities, fusion boundaries, and mapping choices associated with each layer or layer group. This suggests that the main conceptual move behind FFM is not a new closed-form optimization criterion, but a change in the representation of the problem: layer fusion becomes language modeling over discrete mapping decisions (Kao et al., 2022).

4. Architecture, training regime, and one-shot inference

The implementation associated with FFM is DNNFuser, a Transformer-based mapper for layer fusion in DNN accelerators. The paper describes it as the first work, to the best of its knowledge, to propose a one-shot inference-based mapper for this setting, contrasting it with search-based state-of-the-art mapping exploration (Kao et al., 2022).

Training proceeds from a dataset of mapping examples. The method uses pretraining as a generative sequence model and then fine-tuning / adaptation to the specific layer-fusion mapping task and accelerator settings. The rationale for generalization is that a pretrained Transformer can capture structural regularities across many mapping sequences, allowing it to infer reasonable decisions for unseen DNNs, unseen layer combinations, and previously unobserved fusion configurations (Kao et al., 2022).

At inference time, the pipeline is explicitly one-shot:

1. encode the new DNN/hardware instance,
2. feed the encoded sequence into the Transformer,
3. generate mapping tokens sequentially,
4. decode the token sequence into a candidate layer-fusion mapping.

The critical difference from search-based mappers is that DNNFuser avoids runtime enumeration and iterative refinement. The search cost is amortized into training, so deployment reduces to forward-pass generation. This is the operational sense in which the mapper is “fast” (Kao et al., 2022).

5. Empirical claims and performance interpretation

The central empirical claim is that DNNFuser can, within one inference pass, infer solutions with compatible performance to those found by a highly optimized search-based mapper while being 66x–127x faster (Kao et al., 2022). The speedup claim is central because it addresses the runtime overhead that typically accompanies combinatorial map-space exploration.

The paper’s significance lies in its positioning relative to conventional mappers. Existing state-of-the-art mapping explorations are described as search-based, whereas DNNFuser is presented as a learned inference-time substitute. The result is not that exhaustive search has become unnecessary in principle, but that its role can be displaced in practice by a model that has already internalized regularities of good fusion plans (Kao et al., 2022).

A plausible implication is that FFM represents an amortized optimization viewpoint for DNN accelerator mapping: repeated runtime search is replaced by learned priors over good mapping sequences. The paper does not claim guaranteed optimality; rather, it emphasizes compatible quality at much lower mapping latency. That balance between mapping quality and runtime is the most precise meaning of the “fusiest” descriptor in this context.

6. Conceptual boundaries, assumptions, and limitations

A frequent misunderstanding would be to treat FFM as a standalone algorithm separate from DNNFuser. The safer reading is the one given in the paper’s description: FFM is the desired mapper behavior or characterization, whereas DNNFuser is the proposed Transformer-based implementation that realizes it (Kao et al., 2022).

The method also has clear assumptions. It depends on the quality and coverage of training data; if the training set does not include a sufficiently diverse range of fusion patterns or hardware regimes, generalization may weaken. It assumes that layer-fusion mapping can be expressed cleanly as a sequence generation problem, and that the chosen representation preserves the hardware constraints needed for legality. If critical constraints are omitted from the representation, the model may produce suboptimal or invalid fusion plans (Kao et al., 2022).

The paper also places natural boundaries on generalization. Like other pretrained models, DNNFuser may struggle on out-of-distribution accelerators or highly novel DNN structures without fine-tuning. More fundamentally, the approach prioritizes inference speed over exhaustive optimality, so it is a learned approximation rather than a guaranteed optimal solver (Kao et al., 2022).

Taken together, these constraints locate FFM within a broader transition in accelerator CAD and mapping research: from explicit online search toward learned inference-time decision making. In that sense, FFM denotes not merely a fast mapper, but a specific research stance on how inter-layer fusion should be computed—through sequence-model-based inference over combinatorial map-space rather than repeated runtime search.