Land the allocator and AIE codegen changes IRON's stream-dse operators need

[asyms]

What IRON's stream-dse operators (amd/IRON#227) need from stream, on top of main.

1. Allocator plumbing: model expressions registered as named quantities, SolverModel.value, and optional constraint families selected per solve through SolveOptions.families, validated when the options are resolved.
2. A registry of the top-level memory ports of each core (AIE tile DMA channels, ZigZag memory ports, measured per-port bandwidth) and the memory_ports family. With interval and burst off it only reports, which is how IRON runs it: per-port, shared-bandwidth and link activity in the performance view.
3. Transfer targets credited with the tile their consumer core receives.
4. AIE model and codegen fixes found by checking stream's estimates against NPU2:
   • kernel dimensions addressed by operand axis, called once per leading batch index, and a matmul batch axis of one indexed by the node's batch dimension
   • a loop around a re-reading shim transfer chained into one task where its buffer descriptors fit
   • a window an outer loop moves on may be held in one buffer, and each run is charged the fill of the off-chip tensors it holds that way
   • fifos deepened as far as their tile has slack, a broadcast's readers together
   • a transferred tensor placed only where its chosen path delivers it, and a loop of one kept out of a tile's local axes
   • each call linked to the object built for the tile it takes
   • tile DMA channels at 64 bits a cycle, the reconfiguration per column calibrated to IRON's dispatch, and a design whose steady state is one unsplit iteration
   • matmul_PV passed the key length mlir-aie 1.4.4.dev73's kernel reads
5. Dead code removed: the ComputeAllocator module, unused constraint profiles, visualizations, allocator helpers and accessors.

The fast suite passes. With IRON's branch on mlir-aie 1.4.4.dev73, every MHA and SwiGLU stream-dse operator test passes on an NPU2 (85 tests, extensive included).

[github-actions]

Stream AIE Metrics Regression Guard

⚠ 15 cell(s) flagged (total_latency > 0.1% tol): hardware_swiglu[eyeriss_like_dual_core], hardware_swiglu[eyeriss_like_quad_core], hardware_swiglu[eyeriss_like_single_core], hardware_swiglu[fusemax], hardware_swiglu[meta_prototype], hardware_swiglu[simba], hardware_swiglu[simba_small], hardware_swiglu[tpu_like_quad_core], hardware_two_conv[eyeriss_like_dual_core], hardware_two_conv[eyeriss_like_quad_core], hardware_two_conv[fusemax], hardware_two_conv[meta_prototype], hardware_two_conv[simba], hardware_two_conv[simba_small], hardware_two_conv[tpu_like_quad_core]

16 of 16 cells captured

Provenance: baseline 1847bde900d7f4688ec06e9712148aaed915db03 | date 2026-08-18 | Python 3.12.3 | backend ortools_gscip
Note: mip_gap: null — OR-Tools GSCIP
Columns: array fill = per-layer PE-array spatial fill (dataflow quality); MAC eff (e2e) = useful MACs / (chip peak MACs/cycle × total latency), the true fraction of the chip's compute used (incl. idle cores, temporal stalls & transfers).

hardware_swiglu — 8 hardware (⚠ 8 flagged)

seq_len=256, embedding_dim=2048, hidden_dim=8192, bf16; layer-fused tiles seq=16/embedding=128/hidden=32

Hardware	total_latency (base → cur)	Δ%	array fill	MAC eff (e2e)	note
eyeriss_like_dual_core	170918067 → 149684492	⚠ ↓-12.42%	52%	26%
eyeriss_like_quad_core	127074661 → 101581557	⚠ ↓-20.06%	52%	19%
eyeriss_like_single_core	343933021 → 303038604	⚠ ↓-11.89%	52%	25%
fusemax	76808227 → 233766951	⚠ ↑+204.35%	9.3%	0.26%
meta_prototype	110100501 → 101711933	⚠ ↓-7.62%	100%	6.2%
simba	38666519 → 125344507.60	⚠ ↑+224.17%	65%	1.0%
simba_small	123142418 → 106627594	⚠ ↓-13.41%	65%	3.0%
tpu_like_quad_core	103546916 → 101580877	⚠ ↓-1.90%	97%	3.1%

hardware_two_conv — 8 hardware (⚠ 7 flagged)

batch=1, in_ch=8, H=32, W=32, out_ch1=16, out_ch2=32, kernel=3x3, bf16 (generic auto-tiling, not layer-fused)

Hardware	total_latency (base → cur)	Δ%	array fill	MAC eff (e2e)	note
eyeriss_like_dual_core	76675 → 74631	⚠ ↓-2.67%	51%	24%
eyeriss_like_quad_core	41961 → 39917	⚠ ↓-4.87%	51%	22%
eyeriss_like_single_core	114999 → 115003	↑+0.00%	68%	31%
fusemax	186148 → 187612	⚠ ↑+0.79%	1.4%	0.05%	low array utilization
meta_prototype	17385 → 15341	⚠ ↓-11.76%	51%	19%
simba	4316 → 4610.67	⚠ ↑+6.83%	50%	4.9%
simba_small	10750 → 8706	⚠ ↓-19.01%	50%	17%
tpu_like_quad_core	14344 → 12300	⚠ ↓-14.25%	42%	12%

To regenerate baseline: python scripts/analysis/render_metrics_comment.py --update-baseline