Simulator and OpExec Facts

【免费下载链接】cannbot-skills CANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。 项目地址: https://gitcode.com/cann/cannbot-skills

Use this file for simulator behavior one-liners and OpExec call-site gotchas. Use agent/references/code-paths.md and agent/references/simulator-v2.md when you need the full implementation path.

Simulator-behavior one-liners

• OpExec(..., simulator=True | "v2" | "legacy") all route to V2
• simulator="legacy" does not select a separate old runtime
• bar_all() is the only cross-pipe drain in V2; bar_v() / bar_m() are no-ops
• for kernel cycle optimization, the target is the trace makespan (the last instruction/task to finish, i.e. max(ts + dur) over timed trace events), not the sum of all activated-task durations
• wait_vec / wait_cube timeout almost always means the other lane's actor thread crashed silently
• do not run multiple V2 simulator processes concurrently; thread contention can cause silent data corruption
• PyTorch does not support indexing on float8_e5m2 / float8_e4m3fn; view as torch.uint8 before indexing inside @vf
• burst-copy ops (gm_to_ub_pad, ub_to_gm_pad, ub_to_l1_nz) are safe on column-sliced UB views because they use _linear_view_from_pointer

OpExec call-site checklist

• provide shape_bindings={...} when two or more kernel-side scalar dimensions can take the same integer value at runtime
• shape_bindings belongs on the returned callable, not on the OpExec(...) constructor
• format: {tensor_arg_index: [scalar_idx_for_axis_0, scalar_idx_for_axis_1, ...], ...}
• the key is indexed among tensor args only; scalar args are skipped
• use None to keep an axis unbound

Example, for kernel(x:[M,K], y:[N,K], z:[M,N], M, N, K):

shape_bindings={0: [0, 2], 1: [1, 2], 2: [0, 1]}

Implementation:

• easyasc/torchplugin.py:614-668

Real references:

• agent/example/kernels/a5/matmul_rowwise_norm_large_nk.py:137
• agent/example/kernels/a5/vec_cube_vec_scale2_abs_add1_matmul.py:116

Deeper references

• agent/references/simulator-v2.md
• agent/references/code-paths.md
• agent/playbooks/kernel-debugging.md

【免费下载链接】cannbot-skills CANNBot 是面向 CANN 开发的用于提升开发效率的系列智能体，本仓库为其提供可复用的 Skills 模块。 项目地址: https://gitcode.com/cann/cannbot-skills