Task Execution
[toc]
Command Overview
python3 -m ttk {kernel,aclnn,e2e,info,list} [options]
| Subcommand |
Purpose |
kernel |
AscendC kernel compile + NPU execute + precision compare |
aclnn |
aclnn* C API call + precision compare |
e2e |
PyTorch framework API end-to-end test (NPU/MLU/CPU) |
info |
Query local Ascend NPU device info |
list |
Preview test case names from CSV |
python3 -m ttk -v # Show version
python3 -m ttk kernel --help # Show help
Quick Start
# Kernel
python3 -m ttk kernel -i examples/case_store/kernel/mat_mul_v3.csv
python3 -m ttk kernel -i examples/case_store/kernel/add.csv -d
python3 -m ttk kernel -i examples/case_store/kernel/add.csv --co
# ACLNN
python3 -m ttk aclnn -i examples/case_store/aclnn/aclnn_cat.csv
# E2E (auto-selects available backend per configured hardware segment)
python3 -m ttk e2e -i examples/case_store/e2e/torch_add.csv
python3 -m ttk e2e -i examples/case_store/e2e/torch_add.csv --cpu
# Device info & case preview
python3 -m ttk info
python3 -m ttk list -i cases.csv --op add
E2E/ACLNN/Kernel Two-Stage Execution
# Prepare input and CPU golden without target API execution or comparison.
python3 -m ttk e2e -i cases.csv --plugin /path/to/assets \
--no-prof --dump in,golden --dump-format bin \
--manual-data-dirs /data/manual
# Restore prepared data, execute the target API, and compare.
python3 -m ttk e2e -i cases.csv --plugin /path/to/assets \
--manual-data-dirs /data/manual
ACLNN uses the same options; add --plat=<target-soc> when a prepare host cannot
detect the SoC. See Manual-Data Prepare and Replay
for formats, typed-data filename validation, directory defaults, and option constraints.
Kernel uses the same stage selectors while preserving its execution-mode options:
python3 -m ttk kernel -i kernel_cases.csv --plugin /path/to/kernel_assets \
-d=false -c=false -b=release \
--no-prof --dump in,golden --dump-format bin \
--manual-data-dirs /data/kernel_manual
python3 -m ttk kernel -i kernel_cases.csv --plugin /path/to/kernel_assets \
-d=false -c=false -b=release \
--manual-data-dirs /data/kernel_manual
Standalone Kernel --no-prof keeps its existing dry-run behavior. Only the exact
--no-prof --dump in,golden pair selects prepare; --compile-only cannot be used
with either manual-data stage.
Case Selection
| Parameter |
Short |
Description |
Example |
--testcase |
-t |
Filter by case name (comma-separated) |
-t add_01,add_02 |
--testcase-index |
--ti |
Filter by index |
--ti=1,3,5 or --ti=1-10 |
--testcase-count |
--tc |
Randomly pick N cases |
--tc=10 |
--operator |
--op |
Filter by operator name |
--op add,mat_mul_v3 |
--exclude-operator |
--no-op |
Exclude operator name |
--no-op concat_d |
--priority |
|
Filter by priority range |
--priority=1-3 |
--rerun |
|
Rerun failed cases |
--rerun=precision_status |
Device & Parallelism
| Parameter |
Short |
Description |
Default |
--device |
--dev |
Number of devices to use |
All |
--device-whitelist |
|
Device whitelist |
None |
--device-blacklist |
|
Device blacklist |
None |
--process-count |
--pc |
Processes per device |
1 |
--platform |
--plat |
SoC version |
Auto-detect |
--proc-timeout |
|
Per-case timeout (seconds) |
0 (unlimited) |
--limit |
-l |
Memory limit per case (GB); skip if input+output exceeds |
30 |
Precision Control
| Parameter |
Description |
Options |
Default |
--compare |
Comparison method |
close/stat_rel_err/cosine/binary/requant/cross_check |
Spec.tolerance routing (needs --plugin), else stat_rel_err |
--input-dist |
Input distribution |
uniform/normal |
uniform |
--seed |
Random seed (reproducible) |
Integer |
Random |
--golden-mode |
Golden generation mode |
Enable/Disable/Promote |
Enable |
See Result Analysis for comparison method details.
Debug & Diagnostics
| Parameter |
Description |
Example |
--dump |
Dump data: full/in/out/golden |
--dump full |
--dump-format |
Dump format: bin/npy/pt/print |
--dump-format npy |
--dump-on-fail |
Auto-dump all data on precision failure |
--dump-on-fail |
--manual-data-dirs |
Prepare output or ordered replay search roots |
--manual-data-dirs /data/op |
--single-log |
One log file per test case |
--single-log |
--plugin |
External plugin path |
--plugin /path/to/plugin.py |
--validate |
Validate CSV format only (skips compilation, input/golden generation, device execution) |
--validate |
Output
| Parameter |
Short |
Description |
Example |
--output |
-o |
Output result CSV path |
-o results.csv |
--title |
|
Custom output columns |
--title testcase_name,precision_status |
--csv-preserve |
|
Preserve original CSV headers |
--csv-preserve |
Kernel Compilation Control
| Parameter |
Short |
Description |
Default |
--dynamic |
-d |
Dynamic shape compilation; -d false to disable |
On |
--const |
-c |
Static shape compilation |
Off |
--binary |
-b |
Binary mode; -b release for released kernels |
Off |
--compile-only |
--co |
Compile only, skip execution |
Off |
--no-prof |
|
Kernel dry-run; with exact --dump in,golden, select manual-data prepare |
Off |
--compile-opts |
|
Compile options (KEY=VALUE, can be specified multiple times) |
None |
--tiling-run |
--tr |
Tiling run times |
3 |
--reuse-hbm |
|
Each case runs 3 times on NPU by default; reuse same HBM memory to enable L2 Cache |
Off |
--reserve-hbm |
|
Reserve HBM (MB) |
None |
--clear-atomic |
|
Force clear output and workspace before execution |
Off |
--clear-ub / --clear-l1 |
|
Fill UB / L1 with specified value before execution (default: 0) |
Off |
--simt-ub / --simt-stack-dcu |
|
SIMT-mode UB / DCU stack size |
None |
--force-block-dim |
|
Force block_dim value |
None |
E2E-Specific Parameters
E2E mode runs through a unified Backend abstraction (framework_api/backends/); all backends share the same case parsing and precision comparison pipeline. The CPU backend is commonly used as the Golden source. The backend is auto-selected per the configured hardware segment (yaml frameworks.torch.<seg>); --cpu forces the CPU backend.
Two-Stage Selector
--no-prof --dump in,golden prepares input/CPU-golden data without calling the
target API or Kernel. See the two-stage guide above for required combinations.
General Parameters
| Parameter |
Short |
Description |
Default |
--input |
-i |
CSV test case file (required) |
|
--output |
-o |
Output result CSV path |
None |
--seed |
|
Random seed |
Random |
--print |
|
Print summary info |
On |
--no-memory-check |
|
Skip host memory check |
Off |
--proc-no-reuse |
|
New process per case |
Off |
--task-prof |
|
Task-level profiling switch |
On |
--po / --progress-output |
|
Progress output path |
None |
--run |
|
Execution count |
3 (onboard) / 1 (model) |
--warmup |
|
Warmup before profiling |
On |
--npu-timeout |
|
NPU execution timeout (ms) |
Unlimited |
