Skip to content

support.md

Latest commit

 

History

History
48 lines (38 loc) · 3.45 KB

support.md

File metadata and controls

48 lines (38 loc) · 3.45 KB
Raw

Operator and Feature Support

Operators

Operator / feature Current support Notes and limitations
Dense / GEMM Int8/int16, BF16, FP8 Read-only RTP weights; optional bias and fused ReLU; configurable cascade parallelism.
Dynamic MatMul Int8/int16, BF16, FP8 Rank-2 GEMM ABI. A rank-2 RHS may be broadcast across compacted independent LHS axes; batched RHS MatMul with non-broadcast leading axes is rejected.
Elementwise Add Quantized integer Exact-shape inputs for residual and elementwise connections; broadcasting is not supported.
LayerNorm Quantized integer Last-axis normalization using integer mean/variance and reciprocal-square-root approximation. Requires supported static quantization and kernel legality constraints.
Softmax Int8 to uint8/int16 Accurate integer exponential or an opt-in surrogate (needs explicit parameters + QAT). Linear or microtile layout; the exp variant is not yet perf-optimized.
Transpose / Permute Memtile-backed view Permutation of the final two axes only.
Split / Slice Direct or per-slice memtile No Split/Slice kernel. A slice must be an exact union of complete producer-port regions; crossing producer ports, graph-boundary slices, and chained views are not supported.
Concat Direct or per-input memtile No Concat kernel. Each consumer port must read entirely from one input slice; graph-input-backed Concat and chained views are not supported.
Fanout / branching Direct or per-consumer memtile One producer tensor may feed multiple independently planned consumer transport legs.
Constant scale Quantized integer Constant power-of-two output scaling is folded into Dense/MatMul output shifts. Arbitrary constant scaling is not yet supported.
Activation ReLU Fused into Dense; no standalone activation kernel.

Frontends

Frontend Current support Notes and limitations
ONNX Recommended operator-level frontend Supports explicit graphs composed from supported operators and quantized Q/DQ boundaries.
hls4ml Optional MLP-oriented frontend Intended primarily for Dense-style pipelines. Install hls4ml separately when using this path.

Tensor and View Contracts

  • AIE execution buffers remain 2-D. Rank-preserving logical views are compacted only where doing so preserves operator semantics.
  • Dynamic MatMul rejects a non-broadcast batched RHS. Such operations must be lowered into parallel rank-2 MatMul subgraphs or use a future batched-MatMul implementation.
  • Permute supports the final two axes only.
  • Split/Slice and Concat are folded view operations and do not instantiate AIE kernels.
  • Chained folded views are not currently supported.
  • Per-tensor static quantization is the primary supported quantization contract. Per-channel activation quantization is not supported.

Transport

  • Internal transport is realized as either a direct AIE connection or one memory-tile stage.
  • Fanout creates independently planned transport legs for each consumer.
  • Split/Slice and Concat may use direct connections when producer and consumer port regions align exactly; otherwise each legal leg may use a memory tile.
  • Graph boundaries may expose multiple ports for partitioned tensors.
  • Multi-stage relay transport is not implemented. Topologies requiring an additional relay stage fail explicitly.
  • Internal AIE-to-PL-to-AIE bridge points and complete Versal system-link generation are not yet implemented.