xilinx/mlir-aie/aie-model-baseline

Model baseline & numeric oracle (the "before you write any IRON" phase)

This is the first of four phases in porting a model to AIE/NPU:

1. Model baseline & numeric oracle (this skill) — get a validated software

reference and a quantization spec before touching hardware or dataflow.

1. Pre-hardware dataflow simulation (aie-dataflow-presim) — validate the

design (FIFOs, kernels, novel mechanisms) against the oracle in software.

1. Hardware bring-up against reference (aie-hw-bringup) — get a minimal

version running on real silicon, comparing against the oracle at every step.

1. Kernel optimization (aie-kernel-opt) and dataflow optimization

(aie-dataflow-opt) — once it's correct, make it fast.

Skipping this phase doesn't save time — it moves the cost downstream, where it's an order of magnitude more expensive to pay: a scale/shift/rounding bug found here is a five-minute numpy fix; the same bug found during hardware bring-up looks like a mysterious accuracy cliff or a hang, and costs days to isolate.

1. Lock the quantization scheme before designing anything downstream

Pick precision and quantization granularity (per-tensor / per-channel / per-token) before any kernel or dataflow design starts — the scheme determines what the kernels have to do (a fixed power-of-2 shift vs. a runtime-variable one; a per-row reduce_max before requant vs. not), so changing it later means redesigning kernels, not just retuning them.

• If your quantizer claims power-of-2 scales, verify it at design time,

not after kernel-writing has started — walk the quantized graph and check every scale is actually 2^k. If it isn't, either fix the quantizer config or accept the cost now (kernels will need a variable-shift path, which is slower and a real design decision, not a bug to fix later).

• **The right granularity is model-class dependent, and the wrong guess is

invisible in a per-op unit test. A spatial CNN can often tolerate per-tensor activation quantization; an LLM frequently cannot — outlier channels in transformer activations produce a catastrophic accuracy cliff (a measured case: +1300% perplexity) under per-tensor quantization that disappears under per-token/per-row scaling. This only shows up when you validate the full model on real data**, never in an isolated op test with random inputs. If you're quantizing a transformer, budget for a per-row reduce_max before every requant from the start — it's cheap in hardware but expensive to retrofit into kernels and dataflow after the fact.

• Decide now whether activation/weight/KV-cache (if the model is

recurrent/autoregressive) each need their own granularity — they're independent decisions with independent hardware costs.

2. Budget the whole model against a single dispatch before designing anything downstream

Both reference ports this skill family is grounded in (a small CNN classifier, a 1B-parameter LLM) fit entirely within a single NPU dispatch — one compiled design, one set of tiles, no host round-trip mid-model. That's not automatic for a larger model, and getting it wrong looks like everything else in this phase working perfectly right up until dataflow design starts, at which point nothing fits.

Before committing to a design, do a rough budget: total weight bytes at your locked quantization scheme, a rough peak activation footprint, and instruction/program-memory needs, against the NPU configuration's actual ceilings (columns available × per-tile L1, memtile/L2 capacity, program memory per tile, how many distinct kernel bodies can coexist). If it doesn't clearly fit with headroom, decide now — not once dataflow design is underway — where the model splits into multiple dispatches (separate compiled designs, each invoked separately, with intermediate results round-tripping through host/DRAM between them).

Split at a natural boundary the model already has (a layer/block boundary that was going to produce an intermediate tensor anyway) rather than mid-block — a mid-block split adds a host round-trip exactly where none was needed and can force more intermediate state to stay live than the boundary naturally requires.

Once dispatch boundaries are decided, extend the manifest and oracle from this phase to make each dispatch's boundary tensors first-class: the oracle needs to expose the exact intermediate activations at each cut point, because aie-dataflow-presim/aie-hw-bringup will validate each dispatch independently against those tensors — the same way per-block validation works for a single-dispatch design, except these boundaries are real host-mediated hops, not just internal tile boundaries. See aie-dataflow-opt for how the boundary choice itself becomes a performance decision later, not just a feasibility one.

This is the least battle-tested section in this skill family — both source projects were small enough to skip it entirely. Verify the specific ceilings you use against current toolchain/hardware documentation rather than treating any number here as fixed.

3. Export ONNX, then extract a deployment manifest — don't rely on the graph itself as the source of truth

Walking the quantized ONNX graph by hand at kernel-design time, over and over, is slow and error-prone. Instead, do it once: extract every op's quantization parameters (scales, zero points, right-shifts, bias pre-shift/promotion scheme) and its data layout into a single manifest (e.g. a JSON file) up front. This becomes:

• The source of truth that kernel generation and IRON design consume,

instead of everyone re-deriving math from the graph independently.

• A regression artifact — if a re-quantization or re-export changes a

scale, the manifest diff shows exactly what moved.

Record weight layout (e.g. OIYX vs. a packed/tiled format) as a separate field from quantization — they're independent transforms, and keeping them separate lets you test layout round-trips without needing any quantization math, and vice versa.

4. Build a bit-exact numeric oracle, and validate the manifest against it before writing any device code

Write a pure-software reference (numpy or equivalent) that implements the exact arithmetic your kernels will use — the same rounding mode, the same accumulator width/overflow behavior, the same shift-then-clamp order — not just a floating-point approximation. Then validate:

• Manifest extraction against this oracle, bit-exactly, before any IRON

or kernel work starts. If your numpy INT8 conv/matmul doesn't match your extracted scales/shifts bit-for-bit against ONNX Runtime on the same input, the bug is in your understanding of the quantization scheme — find it here, where it costs a numpy print statement, not on hardware where it costs a trace capture and a guessing game.

• **The full model on real data, not just isolated ops with random

inputs.** Random-input unit tests on individual ops are necessary but not sufficient — they systematically miss data-dependent failure modes (the outlier-channel quantization break above is invisible to a per-op random test and only appears running real text/images through the whole model).

This oracle is not throwaway scaffolding — it's the reference every later phase compares against: the software dataflow simulator (phase 2) validates against it, and hardware bring-up (phase 3) validates against it again on real silicon. Time spent making it trustworthy now pays off three times over.

Checklist before moving to phase 2 (aie-dataflow-presim)

• [ ] Quantization scheme chosen and its granularity validated on the full

model with real data (not just random-input op tests)

• [ ] Power-of-2 scale assumption checked, if your kernel design depends on

it

• [ ] Model's total weight/activation/instruction footprint budgeted

against the NPU's single-dispatch ceiling; if it doesn't fit, dispatch boundaries are decided now (ideally at natural layer/block boundaries) and the manifest/oracle expose each boundary's intermediate tensors

• [ ] Deployment manifest extracted (scales/zero-points/shifts/layout per

op) as a standalone artifact, not re-derived from the graph ad hoc

• [ ] Numeric oracle implemented with the exact target arithmetic

(rounding mode, accumulator width, shift order) and validated bit-exactly against the manifest + ONNX Runtime

• [ ] Oracle validated against the full model on real data, not only

per-op random-input tests