Dataset viewer tables: tasks and solutions
Browse filesCo-authored-by: Shuaiwei Huang <112870732+ArtificialRay@users.noreply.github.com>
Co-authored-by: Allen Wu <110312634+wuallenwu@users.noreply.github.com>
Co-authored-by: Weihua Du <73473861+StigLidu@users.noreply.github.com>
- README.md +15 -0
- subset_sme2.json +1 -1
- subset_sve.json +1 -1
- viewer/solutions.parquet +3 -0
- viewer/tasks.parquet +3 -0
README.md
CHANGED
|
@@ -18,6 +18,12 @@ tags:
|
|
| 18 |
pretty_name: KernelShift
|
| 19 |
size_categories:
|
| 20 |
- n<1K
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
---
|
| 22 |
|
| 23 |
# KernelShift
|
|
@@ -69,8 +75,17 @@ workloads/<op_type>/<name>.jsonl one workload per line: axis values and how t
|
|
| 69 |
solutions/<source>/<author>/... kernels, with their sources inlined in the JSON
|
| 70 |
tensors/gemm/<name>/*.npy recorded activations used by the e2e workloads
|
| 71 |
results/e2e-qwen3.5-4b.json end-to-end results for Qwen3.5-4B
|
|
|
|
|
|
|
| 72 |
```
|
| 73 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
Solution authors:
|
| 75 |
|
| 76 |
- `reference-scalar`: scalar code, the correctness reference and the
|
|
|
|
| 18 |
pretty_name: KernelShift
|
| 19 |
size_categories:
|
| 20 |
- n<1K
|
| 21 |
+
configs:
|
| 22 |
+
- config_name: tasks
|
| 23 |
+
data_files: viewer/tasks.parquet
|
| 24 |
+
default: true
|
| 25 |
+
- config_name: solutions
|
| 26 |
+
data_files: viewer/solutions.parquet
|
| 27 |
---
|
| 28 |
|
| 29 |
# KernelShift
|
|
|
|
| 75 |
solutions/<source>/<author>/... kernels, with their sources inlined in the JSON
|
| 76 |
tensors/gemm/<name>/*.npy recorded activations used by the e2e workloads
|
| 77 |
results/e2e-qwen3.5-4b.json end-to-end results for Qwen3.5-4B
|
| 78 |
+
viewer/tasks.parquet one row per definition, for the dataset viewer
|
| 79 |
+
viewer/solutions.parquet one row per solution, for the dataset viewer
|
| 80 |
```
|
| 81 |
|
| 82 |
+
The two `viewer/` tables are generated from the files above so the Hugging Face
|
| 83 |
+
dataset viewer can show them; the benchmark code reads the original files.
|
| 84 |
+
`tasks` has each definition's source, op type, evaluation-set and subset
|
| 85 |
+
membership, workload count, baselines and PyTorch reference (shapes as JSON
|
| 86 |
+
strings). `solutions` has each solution's author, description and main source
|
| 87 |
+
file (`kernel.cpp`, or `binding.cpp` for KleidiAI).
|
| 88 |
+
|
| 89 |
Solution authors:
|
| 90 |
|
| 91 |
- `reference-scalar`: scalar code, the correctness reference and the
|
subset_sme2.json
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
{
|
| 2 |
-
"description": "32-definition dtype-symmetric representative sample of bench-trace/definitions/, restricted to definitions with a real baseline solution compatible with --isa sme2 (mac-m4.metal). Same ncnn/llama.cpp/simd-loop picks as subset_sve.json where their baselines are ISA-agnostic; kleidiai swapped to its sme2 counterpart. loop_010
|
| 3 |
"isa": "sme2",
|
| 4 |
"count": 32,
|
| 5 |
"definitions": [
|
|
|
|
| 1 |
{
|
| 2 |
+
"description": "32-definition dtype-symmetric representative sample of bench-trace/definitions/, restricted to definitions with a real baseline solution compatible with --isa sme2 (mac-m4.metal). Same ncnn/llama.cpp/simd-loop picks as subset_sve.json where their baselines are ISA-agnostic; kleidiai swapped to its sme2 counterpart. loop_010 is left out: it is not in expected_sets.json, so results on it cannot be compared with the full set.",
|
| 3 |
"isa": "sme2",
|
| 4 |
"count": 32,
|
| 5 |
"definitions": [
|
subset_sve.json
CHANGED
|
@@ -1,5 +1,5 @@
|
|
| 1 |
{
|
| 2 |
-
"description": "32-definition dtype-symmetric representative sample of bench-trace/definitions/, restricted to definitions with a real baseline solution compatible with --isa sve. loop_010
|
| 3 |
"isa": "sve",
|
| 4 |
"count": 32,
|
| 5 |
"definitions": [
|
|
|
|
| 1 |
{
|
| 2 |
+
"description": "32-definition dtype-symmetric representative sample of bench-trace/definitions/, restricted to definitions with a real baseline solution compatible with --isa sve. loop_010 is left out: it is not in expected_sets.json, so results on it cannot be compared with the full set.",
|
| 3 |
"isa": "sve",
|
| 4 |
"count": 32,
|
| 5 |
"definitions": [
|
viewer/solutions.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:dea7140c34241731ae9958df5a874868bfd022b2859855cf501507caaa2a8a0a
|
| 3 |
+
size 317978
|
viewer/tasks.parquet
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b947c272bccca05930a9ecbab6635931cbaa59393bf9641cf3832e5b808d0863
|
| 3 |
+
size 39667
|