Add: A5 benchmark cases across graph runtimes

[doraemonmj]

Summary

• Add A5 host-build-graph scene cases for the non-Qwen benchmark workloads.
• Add the actual benchmark_bgemm/Case0 workload to A5 TMR and HBG, using A5-specific fixed-128-tile kernels and scalar topology arguments.
• Reuse each existing A5 TMR orchestration and incore source for the other HBG counterparts.
• Keep the change limited to case migration: no runtime, harness, skill, metric, or A2/A3 changes.

Related to #1727.

Coverage matrix

Benchmark workload	A3 TMR	A3 HBG	A5 TMR	A5 HBG
alternating_matmul_add	Case1	Case1	Case1	Case1 (new)
benchmark_bgemm	Case0	Case0	Case0 (new)	Case0 (new)
paged_attention_unroll	Case1, Case2	Case1, Case2	Case1, Case2	Case1, Case2 (new)
paged_attention_unroll_manual_scope	Case1, Case2	Case1, Case2	Case1, Case2	Case1, Case2 (new)
batch_paged_attention	Case1	Case1	Case1	Case1 (new)
qwen3_14b_decode	StressBatch16Seq3500	×	×	×
spmd_paged_attention	Case1, Case2	×	×	×

The five included workloads have full A3/A5 × TMR/HBG case coverage. × marks a counterpart that is not present in the current tree or this PR:

• qwen3_14b_decode: excluded by request.
• spmd_paged_attention: excluded from this PR and left for follow-up.

A5 BGEMM details

• Preserves the A3 benchmark_bgemm/Case0 shape: 500 matmul tasks, 500 add tasks, tile size 128, incore_loop=4, and grid_k=2.
• Uses A5-specific AIC/AIV kernels because the A2/A3 kernels are not directly reusable on A5.
• Supplies graph-topology values as scalar orchestration arguments. A5 HBG builds the graph on the host, so it must not directly dereference a config tensor's device address.
• The A5 HBG test references the A5 TMR kernel and orchestration sources.

Validation

• A5 hardware (a5new), all NPU access through task-submit --device auto:
  • benchmark_bgemm/Case0 TMR + golden: passed (task_20260813_113257_36868417320)
  • benchmark_bgemm/Case0 HBG + golden: passed (task_20260813_113313_3713652657)
  • Remaining four workloads: default HBG/TMR sweep 4/4 + 4/4; selected manual benchmark cases 3/3 + 3/3.
• A5 lock-free scene compilation: 6/6 classes passed on the rebased tree.
• pre-commit passed: headers, platform literals, clang-format, cpplint, ruff, pyright, and whitespace checks.
• git diff --check passed.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: ef75735f-f731-41c3-910b-353111fe429f

📥 Commits

Reviewing files that changed from the base of the PR and between ab9bdb9 and 0e5fb5d.

📒 Files selected for processing (5)

• examples/a5/host_build_graph/bgemm/test_bgemm.py
• examples/a5/host_build_graph/paged_attention_unroll_manual_scope/test_paged_attention_unroll_manual_scope.py
• tests/st/a5/host_build_graph/alternating_matmul_add/test_alternating_matmul_add.py
• tests/st/a5/host_build_graph/batch_paged_attention/test_batch_paged_attention.py
• tests/st/a5/host_build_graph/paged_attention_unroll/test_paged_attention_unroll.py

---

📝 Walkthrough

Walkthrough

Adds five A5 host-build-graph scene tests. The tests register orchestration and device kernels, define execution cases, generate task arguments, compute golden outputs, and run through SceneTestCase.

Changes

A5 host-build-graph benchmark scenes

Layer / File(s)	Summary
BGEMM scene configuration and validation examples/a5/host_build_graph/bgemm/test_bgemm.py	Adds BGEMM kernel wiring, an A5 case, tiled tensor generation, and a batched matmul golden result.
Alternating matmul-add scene tests/st/a5/host_build_graph/alternating_matmul_add/test_alternating_matmul_add.py	Adds orchestration and compute kernels, default and manual cases, deterministic arguments, and matmul-plus-addition validation.
Paged-attention unroll scenes tests/st/a5/host_build_graph/paged_attention_unroll/test_paged_attention_unroll.py, examples/a5/host_build_graph/paged_attention_unroll_manual_scope/test_paged_attention_unroll_manual_scope.py	Adds standard and manual-scope paged-attention unroll scenes with shared input generation and golden-output computation.
Batch paged-attention scene tests/st/a5/host_build_graph/batch_paged_attention/test_batch_paged_attention.py	Adds batched paged-attention kernel wiring, two A5 bfloat16 cases, task arguments, and golden-output writeback.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Mergeability Score: ⚪ Minimal · up to 0e5fb

This change adds localized A5 host-build-graph benchmark cases without changing kernels, runtime behavior, harnesses, skills, or metrics; no actionable merge-blocking risk remains after normal checks and review.

Possibly related PRs

• hw-native-sys/simpler#1733: Introduces Graph Execution infrastructure used by these A5 scenes.
• hw-native-sys/simpler#1771: Adds a related BGEMM host-build-graph benchmark.
• hw-native-sys/simpler#1784: Adds a related paged-attention unroll implementation.

Sequence Diagram(s)

sequenceDiagram
  participant SceneTestCase
  participant HostBuildGraph
  participant DeviceKernels
  participant GoldenHelper
  SceneTestCase->>HostBuildGraph: register kernels and build task arguments
  HostBuildGraph->>DeviceKernels: execute the configured workload
  SceneTestCase->>GoldenHelper: compute expected tensor outputs
  GoldenHelper-->>SceneTestCase: return golden tensors

Loading

Poem

A rabbit hops through graphs so bright,
With tiled sums packed just right.
Attention flows from Q to V,
Golden answers wait to see.
A5 kernels run and cheer—
“Host-built scenes are hopping here!”

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title clearly identifies the A5 benchmark case additions across graph runtime variants.
Description check	✅ Passed	The description directly explains the A5 host-build-graph benchmark additions, coverage, implementation scope, and validation results.

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[ChaoZheng109]

整体 LGTM 👍 —— 用例迁移边界清晰,所有引用的 kernel/orchestration 源码均可解析,HBG 各用例的 generate_args / compute_golden 与对应 A5 TMR 孪生逐一对齐,A5 BGEMM 改用标量传拓扑参数(HBG host 侧构图不能解引用 config 张量的 device 地址)的设计正确,也比 A3 更干净。

有两个非阻塞的遗留项,想请你确认一下是本 PR 顺手补,还是有意留作后续:

1. benchmark_bgemm 用例缺 a5sim 平台覆盖

TMR 和 HBG 的 bgemm Case0 都只声明了 "platforms": ["a5"],只能在真实硬件上跑,不会进 sim CI 通道。对比:

• A3 的 benchmark_bgemm 跑 ["a2a3sim", "a2a3"];
• 本 PR 同批迁移的 alternating_matmul_add Case0 也带了 ["a5sim", "a5"]。

如果 A5 固定 128-tile 的 kernel 能在 a5sim 下正常跑,建议给 bgemm 补上 a5sim,这样能获得 CI 上的免硬件回归保护(否则这两个用例只有靠人手动在 a5 上跑才能发现回归)。如果是 kernel 目前跑不了 a5sim,麻烦在 PR 里说明一下原因。

2. 新增的 A5 TMR benchmark_bgemm 目录缺 README.md

A3 的 examples/a2a3/tensormap_and_ringbuffer/benchmark_bgemm/ 目录里有 README.md,而新建的 examples/a5/tensormap_and_ringbuffer/benchmark_bgemm/ 没有。属文档 parity 的小缺口,补一份(说明 A5 固定 128-tile、标量传参与 A3 的差异)会方便后来者。

两项都不影响合入,确认后即可。