bcui2 commited on
Commit
c73f69a
·
verified ·
1 Parent(s): 4c9ee3a

Add arXiv citation and report the tolerances applied by the correctness checkers

Browse files

README and croissant.json: BibTeX for arXiv:2611.02752 (Croissant citeAs, version 1.0.1). index/benchmarks.parquet and index/golden_outputs.parquet: replace the catalog placeholders (threshold_raw, spec_atol) with the atol/rtol actually applied by the checkers plus tolerance_source. Input archives, baseline code, and inputs.parquet are unchanged.

README.md CHANGED
@@ -37,7 +37,7 @@ HPC-Bench evaluates large language models as optimizers of computational kernels
37
  This dataset holds everything needed to run the benchmark that is not source code of the evaluation pipeline: the standardized inputs at five scales, the reference outputs used for correctness checking, the exact instrumented baseline programs, and a structured index that ties them together. The evaluation pipeline, the LLM-generated candidates studied in the paper, and the per-trial results live in the GitHub repository.
38
 
39
  - Code and pipeline: https://github.com/Deep-Learning-Profiling-Tools/HPC-Bench
40
- - Paper: TBD
41
 
42
  ## Contents
43
 
@@ -130,7 +130,7 @@ One row per (setting, benchmark): 92 EX1, 92 EX2, 24 EX3.
130
  | Baseline code | `baseline_code_path`, `baseline_code_bytes`, `baseline_code_sha256`, `baseline_code_files`, `baseline_kernel_files`, `source_repo`, `source_commit`, `source_path` |
131
  | Inputs | `input_archive_path`, `input_archive_bytes`, `input_archive_sha256`, `input_sizes`, `n_input_sizes` |
132
  | Reference outputs | `golden_output_sizes`, `n_golden_output_sizes`, `golden_output_bytes` |
133
- | Correctness check | `checker`, `output_type`, `output_dtype`, `output_suffix`, `atol`, `rtol`, `spec_atol`, `threshold_raw`, `comparison_rule` |
134
  | Run configuration | `runtime_args_json` (command-line arguments per input size, from `configs/benchmark_args.json`) |
135
  | Provenance | `upstream_suite`, `upstream_url`, `upstream_reference` |
136
 
@@ -149,7 +149,9 @@ A candidate is correct when the program's output, after the candidate kernel rep
149
  | MachSuite | `machsuite` | `check.data` from the upstream suite | the program compares its own result against `check.data` and prints `Success`/`Fail`; the captured stdout is scanned |
150
  | PARSEC (`canneal`) | `parsec` | text summary | the candidate's `Improvement:` value must be at least the reference value (absolute tolerance 1e-9, relative 1e-6) |
151
 
152
- `atol` is the per-benchmark absolute tolerance from `scripts/benchmark_catalog.py` (`threshold_raw` keeps the catalog value verbatim; where it is not numeric the checker falls back to `spec_atol`), and `rtol`/`spec_atol`/`output_dtype` come from `scripts/checkers/output_specs.py`. `comparison_rule` states the rule in words for every row.
 
 
153
 
154
  Coverage: of the 1,016 (setting, benchmark, size) combinations present in the archives, 1,004 carry reference outputs. The exceptions are `lud` medium (EX1, EX2), `spmv` extra-large (EX1, EX2) and mini (EX2), `susan-c-mibench` medium (EX1, EX2), and all sizes of EX3 `srad_v2`; `haccmk` (EX1, EX2) and EX3 `nn` generate their data inside the program and have no per-size directories, and EX3 `backprop`, `bilateral`, and `gaussian` ship fewer than five scales. `index/inputs.parquet` records this per row.
155
 
@@ -189,7 +191,14 @@ The benchmark programs and input data are derived from PolyBench/C, Rodinia, Par
189
 
190
  ## Citation
191
 
192
- TBD
 
 
 
 
 
 
 
193
 
194
  ## Croissant
195
 
 
37
  This dataset holds everything needed to run the benchmark that is not source code of the evaluation pipeline: the standardized inputs at five scales, the reference outputs used for correctness checking, the exact instrumented baseline programs, and a structured index that ties them together. The evaluation pipeline, the LLM-generated candidates studied in the paper, and the per-trial results live in the GitHub repository.
38
 
39
  - Code and pipeline: https://github.com/Deep-Learning-Profiling-Tools/HPC-Bench
40
+ - Paper: https://arxiv.org/abs/2611.02752
41
 
42
  ## Contents
43
 
 
130
  | Baseline code | `baseline_code_path`, `baseline_code_bytes`, `baseline_code_sha256`, `baseline_code_files`, `baseline_kernel_files`, `source_repo`, `source_commit`, `source_path` |
131
  | Inputs | `input_archive_path`, `input_archive_bytes`, `input_archive_sha256`, `input_sizes`, `n_input_sizes` |
132
  | Reference outputs | `golden_output_sizes`, `n_golden_output_sizes`, `golden_output_bytes` |
133
+ | Correctness check | `checker`, `output_type`, `output_dtype`, `output_suffix`, `atol`, `rtol`, `tolerance_source`, `comparison_rule` |
134
  | Run configuration | `runtime_args_json` (command-line arguments per input size, from `configs/benchmark_args.json`) |
135
  | Provenance | `upstream_suite`, `upstream_url`, `upstream_reference` |
136
 
 
149
  | MachSuite | `machsuite` | `check.data` from the upstream suite | the program compares its own result against `check.data` and prints `Success`/`Fail`; the captured stdout is scanned |
150
  | PARSEC (`canneal`) | `parsec` | text summary | the candidate's `Improvement:` value must be at least the reference value (absolute tolerance 1e-9, relative 1e-6) |
151
 
152
+ `atol` and `rtol` are the tolerances the checker applies when correctness is checked with `scripts/arg.py benchmark check`, and `tolerance_source` says where they are defined: the PolyBench checker uses the per-benchmark threshold in `scripts/benchmark_catalog.py`; the Rodinia, Parboil, Codee, and MiBench checkers use the per-benchmark output specification in `scripts/checkers/output_specs.py`; the PARSEC checker uses constants in `scripts/checkers/parsec.py`. Integer outputs and images are compared exactly (`atol = rtol = 0`), and MachSuite rows carry no tolerance because the program checks itself. `comparison_rule` states the rule in words for every row.
153
+
154
+ Four CUDA-only benchmarks (`dwt2d`, `gaussian`, `huffman`, `hybridsort`) have reference outputs in their archives but no registered output specification, so the released checker framework records their checks as skipped and their candidates are not counted as correct in the released EX3 results; they are marked `checker = unregistered`, as is `nn`, which has no reference output.
155
 
156
  Coverage: of the 1,016 (setting, benchmark, size) combinations present in the archives, 1,004 carry reference outputs. The exceptions are `lud` medium (EX1, EX2), `spmv` extra-large (EX1, EX2) and mini (EX2), `susan-c-mibench` medium (EX1, EX2), and all sizes of EX3 `srad_v2`; `haccmk` (EX1, EX2) and EX3 `nn` generate their data inside the program and have no per-size directories, and EX3 `backprop`, `bilateral`, and `gaussian` ship fewer than five scales. `index/inputs.parquet` records this per row.
157
 
 
191
 
192
  ## Citation
193
 
194
+ ```bibtex
195
+ @article{cui2026hpc,
196
+ title={HPC-Bench: A Comprehensive Benchmark for High Performance Computing Codes},
197
+ author={Cui, Bowen and Yin, Junyu and Ramesh, Tejas and Lim, Leo and Hernandez, Oscar and Zhou, Keren},
198
+ journal={arXiv preprint arXiv:2611.02752},
199
+ year={2026}
200
+ }
201
+ ```
202
 
203
  ## Croissant
204
 
croissant.json CHANGED
@@ -82,9 +82,10 @@
82
  "https://opensource.org/licenses/MIT",
83
  "https://github.com/Deep-Learning-Profiling-Tools/HPC-Bench/blob/master/docs/BENCHMARKS.md"
84
  ],
85
- "citeAs": "HPC-Bench: A Comprehensive Benchmark for High Performance Computing Codes. NeurIPS 2026 Evaluations and Datasets Track. Full citation to be added.",
86
- "version": "1.0.0",
87
  "datePublished": "2026-10-31",
 
88
  "keywords": [
89
  "HPC",
90
  "benchmark",
@@ -177,7 +178,7 @@
177
  "description": "Structured index: benchmarks.",
178
  "contentUrl": "https://huggingface.co/datasets/bcui2/HPC-Bench/resolve/main/index/benchmarks.parquet",
179
  "encodingFormat": "application/x-parquet",
180
- "sha256": "62401b5d517a52052bf043e733697f5f8e4384dca883b83429d170929d6be9af"
181
  },
182
  {
183
  "@type": "cr:FileObject",
@@ -195,7 +196,7 @@
195
  "description": "Structured index: golden_outputs.",
196
  "contentUrl": "https://huggingface.co/datasets/bcui2/HPC-Bench/resolve/main/index/golden_outputs.parquet",
197
  "encodingFormat": "application/x-parquet",
198
- "sha256": "5fa95647a359e57198a8be9c9ceebeb091d198c2348a571b7ec5a5788fde47cf"
199
  }
200
  ],
201
  "recordSet": [
@@ -745,7 +746,7 @@
745
  "@type": "cr:Field",
746
  "@id": "benchmarks/checker",
747
  "name": "checker",
748
- "description": "Correctness checker used for the benchmark",
749
  "dataType": "sc:Text",
750
  "source": {
751
  "fileObject": {
@@ -801,26 +802,11 @@
801
  }
802
  }
803
  },
804
- {
805
- "@type": "cr:Field",
806
- "@id": "benchmarks/rtol",
807
- "name": "rtol",
808
- "description": "Relative tolerance from the output specification",
809
- "dataType": "sc:Float",
810
- "source": {
811
- "fileObject": {
812
- "@id": "benchmarks-parquet"
813
- },
814
- "extract": {
815
- "column": "rtol"
816
- }
817
- }
818
- },
819
  {
820
  "@type": "cr:Field",
821
  "@id": "benchmarks/atol",
822
  "name": "atol",
823
- "description": "Absolute tolerance from the benchmark catalog",
824
  "dataType": "sc:Float",
825
  "source": {
826
  "fileObject": {
@@ -833,31 +819,31 @@
833
  },
834
  {
835
  "@type": "cr:Field",
836
- "@id": "benchmarks/spec_atol",
837
- "name": "spec_atol",
838
- "description": "spec_atol",
839
  "dataType": "sc:Float",
840
  "source": {
841
  "fileObject": {
842
  "@id": "benchmarks-parquet"
843
  },
844
  "extract": {
845
- "column": "spec_atol"
846
  }
847
  }
848
  },
849
  {
850
  "@type": "cr:Field",
851
- "@id": "benchmarks/threshold_raw",
852
- "name": "threshold_raw",
853
- "description": "threshold_raw",
854
  "dataType": "sc:Text",
855
  "source": {
856
  "fileObject": {
857
  "@id": "benchmarks-parquet"
858
  },
859
  "extract": {
860
- "column": "threshold_raw"
861
  }
862
  }
863
  },
@@ -1282,7 +1268,7 @@
1282
  "@type": "cr:Field",
1283
  "@id": "golden_outputs/checker",
1284
  "name": "checker",
1285
- "description": "Correctness checker used for the benchmark",
1286
  "dataType": "sc:Text",
1287
  "source": {
1288
  "fileObject": {
@@ -1323,11 +1309,26 @@
1323
  }
1324
  }
1325
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1326
  {
1327
  "@type": "cr:Field",
1328
  "@id": "golden_outputs/rtol",
1329
  "name": "rtol",
1330
- "description": "Relative tolerance from the output specification",
1331
  "dataType": "sc:Float",
1332
  "source": {
1333
  "fileObject": {
@@ -1340,16 +1341,16 @@
1340
  },
1341
  {
1342
  "@type": "cr:Field",
1343
- "@id": "golden_outputs/atol",
1344
- "name": "atol",
1345
- "description": "Absolute tolerance from the benchmark catalog",
1346
- "dataType": "sc:Float",
1347
  "source": {
1348
  "fileObject": {
1349
  "@id": "golden_outputs-parquet"
1350
  },
1351
  "extract": {
1352
- "column": "atol"
1353
  }
1354
  }
1355
  },
 
82
  "https://opensource.org/licenses/MIT",
83
  "https://github.com/Deep-Learning-Profiling-Tools/HPC-Bench/blob/master/docs/BENCHMARKS.md"
84
  ],
85
+ "citeAs": "@article{cui2026hpc,\n title={HPC-Bench: A Comprehensive Benchmark for High Performance Computing Codes},\n author={Cui, Bowen and Yin, Junyu and Ramesh, Tejas and Lim, Leo and Hernandez, Oscar and Zhou, Keren},\n journal={arXiv preprint arXiv:2611.02752},\n year={2026}\n}",
86
+ "version": "1.0.1",
87
  "datePublished": "2026-10-31",
88
+ "dateModified": "2026-11-06",
89
  "keywords": [
90
  "HPC",
91
  "benchmark",
 
178
  "description": "Structured index: benchmarks.",
179
  "contentUrl": "https://huggingface.co/datasets/bcui2/HPC-Bench/resolve/main/index/benchmarks.parquet",
180
  "encodingFormat": "application/x-parquet",
181
+ "sha256": "a7373d4e1c11f661b8f0f74b655d294af49609c9e28c6a94e6c2ea5308f5e464"
182
  },
183
  {
184
  "@type": "cr:FileObject",
 
196
  "description": "Structured index: golden_outputs.",
197
  "contentUrl": "https://huggingface.co/datasets/bcui2/HPC-Bench/resolve/main/index/golden_outputs.parquet",
198
  "encodingFormat": "application/x-parquet",
199
+ "sha256": "17aefc7bc8a3026cda55d817204f5ffdfce6f48f68e25f67bea6f14e1a8c19d3"
200
  }
201
  ],
202
  "recordSet": [
 
746
  "@type": "cr:Field",
747
  "@id": "benchmarks/checker",
748
  "name": "checker",
749
+ "description": "Correctness checker used for the benchmark (unregistered: no output specification in the released checker framework)",
750
  "dataType": "sc:Text",
751
  "source": {
752
  "fileObject": {
 
802
  }
803
  }
804
  },
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
805
  {
806
  "@type": "cr:Field",
807
  "@id": "benchmarks/atol",
808
  "name": "atol",
809
+ "description": "Absolute tolerance applied by the correctness checker",
810
  "dataType": "sc:Float",
811
  "source": {
812
  "fileObject": {
 
819
  },
820
  {
821
  "@type": "cr:Field",
822
+ "@id": "benchmarks/rtol",
823
+ "name": "rtol",
824
+ "description": "Relative tolerance applied by the correctness checker",
825
  "dataType": "sc:Float",
826
  "source": {
827
  "fileObject": {
828
  "@id": "benchmarks-parquet"
829
  },
830
  "extract": {
831
+ "column": "rtol"
832
  }
833
  }
834
  },
835
  {
836
  "@type": "cr:Field",
837
+ "@id": "benchmarks/tolerance_source",
838
+ "name": "tolerance_source",
839
+ "description": "Where the applied tolerance is defined (benchmark catalog, output specification, checker constants, or program self-check)",
840
  "dataType": "sc:Text",
841
  "source": {
842
  "fileObject": {
843
  "@id": "benchmarks-parquet"
844
  },
845
  "extract": {
846
+ "column": "tolerance_source"
847
  }
848
  }
849
  },
 
1268
  "@type": "cr:Field",
1269
  "@id": "golden_outputs/checker",
1270
  "name": "checker",
1271
+ "description": "Correctness checker used for the benchmark (unregistered: no output specification in the released checker framework)",
1272
  "dataType": "sc:Text",
1273
  "source": {
1274
  "fileObject": {
 
1309
  }
1310
  }
1311
  },
1312
+ {
1313
+ "@type": "cr:Field",
1314
+ "@id": "golden_outputs/atol",
1315
+ "name": "atol",
1316
+ "description": "Absolute tolerance applied by the correctness checker",
1317
+ "dataType": "sc:Float",
1318
+ "source": {
1319
+ "fileObject": {
1320
+ "@id": "golden_outputs-parquet"
1321
+ },
1322
+ "extract": {
1323
+ "column": "atol"
1324
+ }
1325
+ }
1326
+ },
1327
  {
1328
  "@type": "cr:Field",
1329
  "@id": "golden_outputs/rtol",
1330
  "name": "rtol",
1331
+ "description": "Relative tolerance applied by the correctness checker",
1332
  "dataType": "sc:Float",
1333
  "source": {
1334
  "fileObject": {
 
1341
  },
1342
  {
1343
  "@type": "cr:Field",
1344
+ "@id": "golden_outputs/tolerance_source",
1345
+ "name": "tolerance_source",
1346
+ "description": "Where the applied tolerance is defined (benchmark catalog, output specification, checker constants, or program self-check)",
1347
+ "dataType": "sc:Text",
1348
  "source": {
1349
  "fileObject": {
1350
  "@id": "golden_outputs-parquet"
1351
  },
1352
  "extract": {
1353
+ "column": "tolerance_source"
1354
  }
1355
  }
1356
  },
index/benchmarks.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:62401b5d517a52052bf043e733697f5f8e4384dca883b83429d170929d6be9af
3
- size 93558
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a7373d4e1c11f661b8f0f74b655d294af49609c9e28c6a94e6c2ea5308f5e464
3
+ size 93598
index/golden_outputs.parquet CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:5fa95647a359e57198a8be9c9ceebeb091d198c2348a571b7ec5a5788fde47cf
3
- size 19585
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:17aefc7bc8a3026cda55d817204f5ffdfce6f48f68e25f67bea6f14e1a8c19d3
3
+ size 21090