Add baseline source archives, Parquet indexes, dataset card, and Croissant metadata (NeurIPS 2026 release)
Browse filesAdds baseline_code/ (208 instrumented baseline source archives + build_infrastructure.zip + manifest.csv), index/{benchmarks,inputs,golden_outputs}.parquet, croissant.json, and a new README.md. The 207 input archives and the root manifest.csv are unchanged.
This view is limited to 50 files because it contains too many changes. See raw diff
- README.md +170 -20
- baseline_code/EX1/2mm.zip +3 -0
- baseline_code/EX1/3mm.zip +3 -0
- baseline_code/EX1/adi.zip +3 -0
- baseline_code/EX1/aes.zip +3 -0
- baseline_code/EX1/atax.zip +3 -0
- baseline_code/EX1/atmux.zip +3 -0
- baseline_code/EX1/b+tree.zip +3 -0
- baseline_code/EX1/backprop.zip +3 -0
- baseline_code/EX1/basicmath-mibench.zip +3 -0
- baseline_code/EX1/bfs-bulk-mach.zip +3 -0
- baseline_code/EX1/bfs-queue-mach.zip +3 -0
- baseline_code/EX1/bfs.zip +3 -0
- baseline_code/EX1/bicg.zip +3 -0
- baseline_code/EX1/bilateral.zip +3 -0
- baseline_code/EX1/bitcount-mibench.zip +3 -0
- baseline_code/EX1/canneal-parsec.zip +3 -0
- baseline_code/EX1/cfd.zip +3 -0
- baseline_code/EX1/cholesky.zip +3 -0
- baseline_code/EX1/correlation.zip +3 -0
- baseline_code/EX1/coulomb.zip +3 -0
- baseline_code/EX1/covariance.zip +3 -0
- baseline_code/EX1/cutcp.zip +3 -0
- baseline_code/EX1/deriche.zip +3 -0
- baseline_code/EX1/dijkstra-mibench.zip +3 -0
- baseline_code/EX1/doitgen.zip +3 -0
- baseline_code/EX1/durbin.zip +3 -0
- baseline_code/EX1/fdtd-2d.zip +3 -0
- baseline_code/EX1/fft-strided-mach.zip +3 -0
- baseline_code/EX1/fft-transpose-mach.zip +3 -0
- baseline_code/EX1/floyd-warshall.zip +3 -0
- baseline_code/EX1/gemm-blocked-mach.zip +3 -0
- baseline_code/EX1/gemm-ncubed-mach.zip +3 -0
- baseline_code/EX1/gemm.zip +3 -0
- baseline_code/EX1/gemver.zip +3 -0
- baseline_code/EX1/gesummv.zip +3 -0
- baseline_code/EX1/gramschmidt.zip +3 -0
- baseline_code/EX1/haccmk.zip +3 -0
- baseline_code/EX1/heartwall.zip +3 -0
- baseline_code/EX1/heat-3d.zip +3 -0
- baseline_code/EX1/histo.zip +3 -0
- baseline_code/EX1/hotspot.zip +3 -0
- baseline_code/EX1/hotspot3D.zip +3 -0
- baseline_code/EX1/jacobi-1d.zip +3 -0
- baseline_code/EX1/jacobi-2d.zip +3 -0
- baseline_code/EX1/kmeans.zip +3 -0
- baseline_code/EX1/kmp-mach.zip +3 -0
- baseline_code/EX1/lavaMD.zip +3 -0
- baseline_code/EX1/lbm.zip +3 -0
- baseline_code/EX1/leukocyte.zip +3 -0
README.md
CHANGED
|
@@ -1,46 +1,196 @@
|
|
| 1 |
---
|
| 2 |
-
pretty_name: HPC-Bench
|
| 3 |
license: other
|
| 4 |
-
license_name: upstream-benchmark-licenses
|
|
|
|
|
|
|
|
|
|
| 5 |
tags:
|
| 6 |
- hpc
|
| 7 |
- benchmark
|
| 8 |
- code-optimization
|
| 9 |
- llm-evaluation
|
|
|
|
| 10 |
- openmp
|
| 11 |
- cuda
|
|
|
|
| 12 |
size_categories:
|
| 13 |
-
- n<
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
---
|
| 15 |
|
| 16 |
-
# HPC-Bench
|
| 17 |
|
| 18 |
-
|
| 19 |
|
| 20 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 21 |
|
| 22 |
## Contents
|
| 23 |
|
| 24 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
-
|
| 27 |
-
|:---|:---|---:|
|
| 28 |
-
| `input_data/EX1/<benchmark>.zip` | Serial CPU optimization | 92 |
|
| 29 |
-
| `input_data/EX2/<benchmark>.zip` | OpenMP CPU parallelization | 92 |
|
| 30 |
-
| `input_data/EX3/<benchmark>.zip` | CUDA GPU optimization | 23 |
|
| 31 |
|
| 32 |
-
`
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
-
##
|
| 35 |
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
```bash
|
| 39 |
-
python
|
| 40 |
-
python tools/fetch_input_data.py --ex EX1 # one setting
|
| 41 |
-
python tools/fetch_input_data.py --benchmarks 2mm,bfs # selected benchmarks
|
| 42 |
```
|
| 43 |
|
| 44 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
-
|
|
|
|
| 1 |
---
|
| 2 |
+
pretty_name: HPC-Bench
|
| 3 |
license: other
|
| 4 |
+
license_name: mixed-upstream-benchmark-licenses
|
| 5 |
+
license_link: https://github.com/Deep-Learning-Profiling-Tools/HPC-Bench/blob/master/docs/BENCHMARKS.md
|
| 6 |
+
language:
|
| 7 |
+
- en
|
| 8 |
tags:
|
| 9 |
- hpc
|
| 10 |
- benchmark
|
| 11 |
- code-optimization
|
| 12 |
- llm-evaluation
|
| 13 |
+
- performance
|
| 14 |
- openmp
|
| 15 |
- cuda
|
| 16 |
+
- c
|
| 17 |
size_categories:
|
| 18 |
+
- 1K<n<10K
|
| 19 |
+
configs:
|
| 20 |
+
- config_name: benchmarks
|
| 21 |
+
default: true
|
| 22 |
+
data_files: index/benchmarks.parquet
|
| 23 |
+
- config_name: inputs
|
| 24 |
+
data_files: index/inputs.parquet
|
| 25 |
+
- config_name: golden_outputs
|
| 26 |
+
data_files: index/golden_outputs.parquet
|
| 27 |
+
- config_name: input_archives
|
| 28 |
+
data_files: manifest.csv
|
| 29 |
---
|
| 30 |
|
| 31 |
+
# HPC-Bench
|
| 32 |
|
| 33 |
+
Benchmark data for **HPC-Bench: A Comprehensive Benchmark for High Performance Computing Codes** (NeurIPS 2026 Evaluations and Datasets Track).
|
| 34 |
|
| 35 |
+
HPC-Bench evaluates large language models as optimizers of computational kernels embedded in complete HPC programs. Each task is a profiler-identified kernel region of a real benchmark program; a candidate replaces only that region, and the whole program is compiled, run on standardized inputs, checked for correctness against a reference output, and timed. The suite covers 92 CPU workloads evaluated under serial (EX1) and OpenMP (EX2) optimization and 22 Rodinia-derived CUDA workloads evaluated under GPU optimization (EX3), spanning 16 computational motifs.
|
| 36 |
+
|
| 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 |
|
| 44 |
+
| Component | Path | Size | Notes |
|
| 45 |
+
|:---|:---|---:|:---|
|
| 46 |
+
| Input data, 5 scales per benchmark | `input_data/{EX1,EX2,EX3}/<benchmark>.zip` | 207 archives, 37.7 GB | One archive per setting and benchmark; also contains the reference outputs |
|
| 47 |
+
| Reference (golden) outputs | inside the input archives, `input_data/<size>/output/` | 1,023 files, 78.9 GB uncompressed | Indexed in `index/golden_outputs.parquet`; not duplicated |
|
| 48 |
+
| Instrumented baseline source | `baseline_code/{EX1,EX2,EX3}/<benchmark>.zip` | 208 archives, 35 MB | Exact baseline programs used in the experiments; no LLM-generated code |
|
| 49 |
+
| Shared build infrastructure | `baseline_code/build_infrastructure.zip` | 190 files | `common.mk`, per-suite Makefile infrastructure, PolyBench utilities |
|
| 50 |
+
| Benchmark index | `index/benchmarks.parquet` | 208 rows | One row per (setting, benchmark) |
|
| 51 |
+
| Input-size index | `index/inputs.parquet` | 1,040 rows | One row per (setting, benchmark, input size) |
|
| 52 |
+
| Reference-output index | `index/golden_outputs.parquet` | 1,023 rows | One row per reference-output file |
|
| 53 |
+
| Archive manifests | `manifest.csv`, `baseline_code/manifest.csv` | | Path, size, and SHA-256 of every archive |
|
| 54 |
+
| Croissant metadata | `croissant.json` | | Croissant 1.1 with NeurIPS RAI fields |
|
| 55 |
+
|
| 56 |
+
Settings: **EX1** serial CPU optimization (`gcc -O3`), **EX2** OpenMP CPU parallelization (`gcc -O3 -fopenmp`, thread count set by `OMP_NUM_THREADS`), **EX3** CUDA GPU optimization (`nvcc`). EX1 and EX2 share the same 92 benchmarks and the same inputs. EX3 contains 24 benchmark directories: the 22 CUDA benchmarks reported in the paper, plus `backprop` and `nw`, which have CUDA sources and inputs but are not part of the paper's CUDA benchmark table (`in_paper_benchmark_set = false`). `nn` generates its input at run time and has no input archive.
|
| 57 |
+
|
| 58 |
+
## Layout
|
| 59 |
+
|
| 60 |
+
```text
|
| 61 |
+
bcui2/HPC-Bench/
|
| 62 |
+
├── README.md
|
| 63 |
+
├── croissant.json
|
| 64 |
+
├── manifest.csv # input archives: ex, benchmark, path, bytes, sha256
|
| 65 |
+
├── input_data/
|
| 66 |
+
│ └── EX1|EX2|EX3/<benchmark>.zip # -> input_data/<size>/{input,output}/...
|
| 67 |
+
├── baseline_code/
|
| 68 |
+
│ ├── manifest.csv
|
| 69 |
+
│ ├── build_infrastructure.zip # common.mk, common_*/, driver_parboil/, utilities/, EX*/Makefile
|
| 70 |
+
│ └── EX1|EX2|EX3/<benchmark>.zip # -> <benchmark>/... (Makefile, sources, EX*_optimized_codes/<baseline kernel>)
|
| 71 |
+
└── index/
|
| 72 |
+
├── benchmarks.parquet
|
| 73 |
+
├── inputs.parquet
|
| 74 |
+
└── golden_outputs.parquet
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
Each input archive unpacks to a single `input_data/` directory with one subdirectory per scale (`mini`, `small`, `medium`, `large`, `extra-large`). Inputs are under `input_data/<size>/input/` (PolyBench kernels generate their inputs in the program and have no input files) and reference outputs under `input_data/<size>/output/`.
|
| 78 |
+
|
| 79 |
+
## Quick start
|
| 80 |
+
|
| 81 |
+
### Inputs and reference outputs
|
| 82 |
+
|
| 83 |
+
Clone the GitHub repository and let its fetch script download, verify, and extract the archives into `EX*/<benchmark>/input_data/`:
|
| 84 |
+
|
| 85 |
+
```bash
|
| 86 |
+
git clone https://github.com/Deep-Learning-Profiling-Tools/HPC-Bench.git
|
| 87 |
+
cd HPC-Bench
|
| 88 |
+
python tools/fetch_input_data.py # everything (~38 GB zipped)
|
| 89 |
+
python tools/fetch_input_data.py --ex EX1 --benchmarks 2mm,bfs
|
| 90 |
+
python tools/fetch_input_data.py --list
|
| 91 |
+
```
|
| 92 |
+
|
| 93 |
+
The script reads `manifest.csv` from this dataset, checks each archive's SHA-256, and extracts it in place. Reference outputs arrive with the inputs.
|
| 94 |
|
| 95 |
+
### Index
|
|
|
|
|
|
|
|
|
|
|
|
|
| 96 |
|
| 97 |
+
```python
|
| 98 |
+
import pandas as pd
|
| 99 |
+
b = pd.read_parquet("hf://datasets/bcui2/HPC-Bench/index/benchmarks.parquet")
|
| 100 |
+
b[b.setting == "EX1"][["benchmark", "suite", "motif_abbrev", "structural_complexity", "checker", "atol", "rtol"]]
|
| 101 |
+
```
|
| 102 |
+
|
| 103 |
+
or with `datasets`: `load_dataset("bcui2/HPC-Bench", "benchmarks")` (also `"inputs"`, `"golden_outputs"`, `"input_archives"`).
|
| 104 |
|
| 105 |
+
### Baseline source code
|
| 106 |
|
| 107 |
+
```python
|
| 108 |
+
from huggingface_hub import hf_hub_download
|
| 109 |
+
import zipfile
|
| 110 |
+
for f in ["baseline_code/build_infrastructure.zip", "baseline_code/EX1/2mm.zip"]:
|
| 111 |
+
p = hf_hub_download("bcui2/HPC-Bench", f, repo_type="dataset")
|
| 112 |
+
zipfile.ZipFile(p).extractall("EX1" if "/EX1/" in f else ".")
|
| 113 |
+
```
|
| 114 |
+
|
| 115 |
+
`baseline_code/EX1/2mm.zip` unpacks to `2mm/...`, so extracting it inside `EX1/` recreates `EX1/2mm/` as it is in the GitHub repository (minus the LLM candidates). `build_infrastructure.zip` unpacks at the repository root. The benchmark Makefiles include the shared infrastructure through relative paths, so keep this layout. Building and running are driven by `scripts/arg.py` in the GitHub repository:
|
| 116 |
|
| 117 |
```bash
|
| 118 |
+
python scripts/arg.py benchmark build --root . --benchmarks 2mm --ex-versions EX1
|
|
|
|
|
|
|
| 119 |
```
|
| 120 |
|
| 121 |
+
## Benchmark index (`index/benchmarks.parquet`)
|
| 122 |
+
|
| 123 |
+
One row per (setting, benchmark): 92 EX1, 92 EX2, 24 EX3.
|
| 124 |
+
|
| 125 |
+
| Column group | Columns |
|
| 126 |
+
|:---|:---|
|
| 127 |
+
| Identity | `benchmark`, `setting` (EX1/EX2/EX3), `setting_name`, `language` (C, C++, CUDA), `suite`, `suite_name` |
|
| 128 |
+
| Classification | `motif`, `motif_abbrev` (16 motifs), `structural_complexity` (d1–d4), `complexity_loc`, `complexity_functions`, `complexity_branches`, `complexity_loop_depth`, `complexity_score`, `in_paper_benchmark_set` |
|
| 129 |
+
| Annotated region | `annotated_target_file`, `annotated_start_line`, `annotated_end_line` (the kernel region the model may rewrite; `annotated_regions_json` keeps multi-region cases verbatim) |
|
| 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 |
+
|
| 137 |
+
`structural_complexity` is the paper's static complexity level computed from the annotated region (branch count, lines of code, loop depth, function count). The raw metrics are given for the CPU benchmarks; for EX3 the level comes from the paper's CUDA benchmark table and the raw metrics are not released.
|
| 138 |
+
|
| 139 |
+
`index/inputs.parquet` has one row per (setting, benchmark, size) with `size_present`, `input_files`, `input_bytes`, `golden_output_available`, `golden_output_files` (paths inside the archive), `golden_output_bytes`, and `runtime_args`. `index/golden_outputs.parquet` has one row per reference-output file with its `path_in_archive`, `bytes`, and the comparison parameters of its benchmark.
|
| 140 |
+
|
| 141 |
+
## Reference outputs and correctness checking
|
| 142 |
+
|
| 143 |
+
A candidate is correct when the program's output, after the candidate kernel replaces the annotated region, matches the reference output for the same input under the suite's comparison rule. The reference output for each input is the output of the instrumented baseline program on that input, stored at `input_data/<size>/output/` inside the input archive; MachSuite ships its own `check.data` reference that the program compares against itself. Both are consumed in place by the checkers in `scripts/checkers/` of the GitHub repository.
|
| 144 |
+
|
| 145 |
+
| Suite | Checker | Reference | Comparison rule |
|
| 146 |
+
|:---|:---|:---|:---|
|
| 147 |
+
| PolyBench | `polybench` | text dump of the output arrays | numeric tokens compared element-wise, `numpy.isclose(atol=atol, rtol=0)`; `atol` is 0.002 (0.0005 for one benchmark) |
|
| 148 |
+
| Rodinia, Parboil, Codee, MiBench | suite checker | text or binary file (`output_type`, `output_dtype`, `output_suffix`) | shapes must match; floating-point data `numpy.allclose(atol=atol, rtol=rtol)`, integer data and PGM images exactly equal |
|
| 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 |
+
|
| 156 |
+
To regenerate a reference output, build the baseline with `scripts/arg.py benchmark build` as above, run the baseline executable for that size with the arguments in `runtime_args`, and capture its output (stdout for text outputs, or the file the program writes) to `input_data/<size>/output/`. The baseline source, compiler flags, and inputs released here are the ones used for the published references; a different compiler or architecture can change floating-point results within the stated tolerances.
|
| 157 |
+
|
| 158 |
+
## Baseline source code
|
| 159 |
+
|
| 160 |
+
`baseline_code/<EX>/<benchmark>.zip` contains the benchmark directory from the GitHub repository at commit `7dc9824ba10249babe9c898b7591415b79445e22`, restricted to the baseline: Makefile, program sources and headers, data generators, and the baseline kernel file(s) in `<EX>_optimized_codes/` (`baseline_kernel_files` in the index). All LLM-generated candidate files (`*_<model>_v<N>.*`, 6,464 files) and all input data are excluded. The kernels carry the timing instrumentation used for the paper's kernel-time measurements; the annotated optimization region is given by `annotated_target_file` and the line range in the index. Every archive is listed with its SHA-256 in `baseline_code/manifest.csv`.
|
| 161 |
+
|
| 162 |
+
## Benchmark composition
|
| 163 |
+
|
| 164 |
+
| Suite | EX1/EX2 workloads | Upstream |
|
| 165 |
+
|:---|---:|:---|
|
| 166 |
+
| PolyBench/C | 30 | https://www.cs.colostate.edu/~pouchet/software/polybench/ |
|
| 167 |
+
| Rodinia | 19 (+ CUDA set) | https://www.cs.virginia.edu/rodinia/ |
|
| 168 |
+
| MachSuite | 18 | https://github.com/breagen/MachSuite |
|
| 169 |
+
| Parboil | 10 | http://impact.crhc.illinois.edu/parboil/parboil.aspx |
|
| 170 |
+
| MiBench | 9 | https://vhosts.eecs.umich.edu/mibench/ |
|
| 171 |
+
| Codee | 5 | https://www.codee.com/ |
|
| 172 |
+
| PARSEC | 1 | https://parsec.cs.princeton.edu/ |
|
| 173 |
+
|
| 174 |
+
Motifs (EX1/EX2 counts): Dense Linear Algebra 22, Stencil / Scientific Computing 13, Image and Video Processing 7, Graph Algorithms 6, N-body Methods 6, Dynamic Programming 5, Sorting / Search 5, Clustering / Machine Learning 4, Medical / Scientific Computing 4, Statistical Computations 4, Utility Benchmarks 4, Monte Carlo Methods 3, Sparse Linear Algebra 3, Cryptography / Encoding 2, Computational Fluid Dynamics 2, Spectral Methods 2. Structural complexity: 61 d1, 17 d2, 9 d3, 5 d4.
|
| 175 |
+
|
| 176 |
+
## Intended use and limitations
|
| 177 |
+
|
| 178 |
+
HPC-Bench is intended for evaluating code-generating models on correctness-preserving performance optimization of HPC kernels under serial, OpenMP, and CUDA settings, for studying the failure modes of such optimizations, and for testing compilers, autotuners, and program-optimization tools against fixed inputs and references. It measures kernel-level optimization of profiled regions, not end-to-end application performance or general software-engineering ability.
|
| 179 |
+
|
| 180 |
+
The benchmarks are kernels embedded in complete but moderately sized programs from academic suites; libraries such as BLAS or MPI are not exercised, and all programs are C, C++, or CUDA. Coverage is skewed toward dense linear algebra and stencil kernels and toward structurally simple regions (61 of 92 CPU workloads are d1); sparse linear algebra is covered by three SpMV kernels. The reference outputs and tolerances were established with GCC 14.2 and CUDA 12.4 at `-O3` on an AMD EPYC 7543 / NVIDIA A100 system; other compilers or architectures can change floating-point results within the stated tolerances. The data contain no personal or sensitive information; inputs are synthetic or standard scientific datasets.
|
| 181 |
+
|
| 182 |
+
## Reproducibility
|
| 183 |
+
|
| 184 |
+
The published evaluation runs entirely from this dataset plus the GitHub repository: `tools/fetch_input_data.py` places the inputs and reference outputs, `baseline_code/` provides the exact baseline programs at a fixed commit, the index gives the annotated region, run arguments, and comparison rule for every benchmark, and `scripts/arg.py` in the repository builds, runs, checks, and times candidates. The LLM candidates and per-trial results behind the paper's figures are in the repository (`EX*/<benchmark>/EX*_optimized_codes/`, `final_data/`), with a figure-by-figure guide in `docs/REPRODUCING_PAPER.md`.
|
| 185 |
+
|
| 186 |
+
## Licensing and attribution
|
| 187 |
+
|
| 188 |
+
The benchmark programs and input data are derived from PolyBench/C, Rodinia, Parboil, MachSuite, MiBench, PARSEC, and Codee and remain subject to the licenses and attribution requirements of those suites; the upstream references are given per row in the index (`upstream_suite`, `upstream_url`, `upstream_reference`) and in the GitHub repository's `docs/BENCHMARKS.md`. The Codee kernels (`atmux`, `coulomb`, `haccmk`, `matmul`, `pi`) are redistributed with Codee's explicit permission, which extends to continued redistribution after the conclusion of the research collaboration in which they were obtained. The index files, manifests, metadata, and the HPC-Bench instrumentation and tooling are released under the MIT License by the HPC-Bench authors. Users who redistribute benchmark programs or inputs should check the license of the originating suite.
|
| 189 |
+
|
| 190 |
+
## Citation
|
| 191 |
+
|
| 192 |
+
TBD
|
| 193 |
+
|
| 194 |
+
## Croissant
|
| 195 |
|
| 196 |
+
`croissant.json` describes this dataset in Croissant 1.1, including the Responsible AI fields required by the NeurIPS 2026 Evaluations and Datasets Track, and was validated with `mlcroissant`.
|
baseline_code/EX1/2mm.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b0f247b828475da6d9207f63ee7fd7c0d3eb935f79db4585a1fce9559b8190c5
|
| 3 |
+
size 9599
|
baseline_code/EX1/3mm.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:cbf4a76556a79975cd28eb31e50379dd32f48af7a2ae3623f8c016f0c8f04e8c
|
| 3 |
+
size 9194
|
baseline_code/EX1/adi.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f6f38317cb87e3fbc17c2e588f9a26562a09405e077bb6ae94ce753f036a00bc
|
| 3 |
+
size 9759
|
baseline_code/EX1/aes.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:52f818e39efa05b704e0d91f6f52f1f70e9c7281dc1dd9d365514fb2c9af82c1
|
| 3 |
+
size 10954
|
baseline_code/EX1/atax.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d39e1e30ade59be187aadb050651ea259f23792c04168ce652438ab155f10120
|
| 3 |
+
size 9072
|
baseline_code/EX1/atmux.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d8a3ad025cde8e5424b45fbd60f9bb6da2964b91c24c37a6fd76123fc4ed089e
|
| 3 |
+
size 7589
|
baseline_code/EX1/b+tree.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a0ff2d4aa6db8cc00feaaf54d9cc72387e81bc6df0dcfeda0379f7493d72534e
|
| 3 |
+
size 30102
|
baseline_code/EX1/backprop.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:17fd76965c793e61270a41651056580166b6561aff0f5d1bea769e049be990e8
|
| 3 |
+
size 8591
|
baseline_code/EX1/basicmath-mibench.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:77d1c4d08c3651374067abb03d6512fd35b50861fbbb535c8cbd08941c6ce617
|
| 3 |
+
size 12477
|
baseline_code/EX1/bfs-bulk-mach.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:44ea65ce4436e3205fac5b2a0195cf494063fd09f28e9e0b51c0ecf19a18fff4
|
| 3 |
+
size 10723
|
baseline_code/EX1/bfs-queue-mach.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f3869a8fcff97a0760e2af3e0feaa5f142c4f2948bc4c184d38041475148581e
|
| 3 |
+
size 8215
|
baseline_code/EX1/bfs.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1dfe114b9d9730d7944ac1d99eaabdebea5614e24a5806d35385997dfa186429
|
| 3 |
+
size 2659
|
baseline_code/EX1/bicg.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3cf08a736541405e91add9bd4aa8c6a8942feb10721139f40891c0266a814cd2
|
| 3 |
+
size 6710
|
baseline_code/EX1/bilateral.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:be54c88dd0c497d7b0805d35a66c8f86c8fbf2828a0e4eca7d5e6b69ccec3be3
|
| 3 |
+
size 11153
|
baseline_code/EX1/bitcount-mibench.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8da2f820c5b0c380de9490152001d60949ca5c603d3ff1903dc8ff36f196b39b
|
| 3 |
+
size 12165
|
baseline_code/EX1/canneal-parsec.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:65b450db484fb08b91f769e30a3ed87768d9b9bb5b02fd04b9d198f1132ee265
|
| 3 |
+
size 76566
|
baseline_code/EX1/cfd.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:601fd6e1db431fb0fe288f27a991daa3bb879737cd8d4dc0b27c812d90fa91ae
|
| 3 |
+
size 22278
|
baseline_code/EX1/cholesky.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fe70619ca642ae72b34cb8216681fe49c334f69e818b8d620a9c1a317b7e1151
|
| 3 |
+
size 5201
|
baseline_code/EX1/correlation.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3418320464f91a69da271b0035160b79ac9228f989584c80758a99e35969337e
|
| 3 |
+
size 8125
|
baseline_code/EX1/coulomb.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:02392a215e608b5229d4d3ecc1166ccc93b3566becc160b28570ff3f5d8971e3
|
| 3 |
+
size 6266
|
baseline_code/EX1/covariance.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:18e70e029bca28acc9f9bfbb0a3ffe5e8757573440cb70a116ace214cfea00e0
|
| 3 |
+
size 7816
|
baseline_code/EX1/cutcp.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d6d1dfc47f7f965043946c77af881491341d59128b87cf28b847e9a8fedb8309
|
| 3 |
+
size 19261
|
baseline_code/EX1/deriche.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:48de51363036ba066415e5214bf908cdac0507a8541e5e6321235e3bb7fbe6ce
|
| 3 |
+
size 8338
|
baseline_code/EX1/dijkstra-mibench.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e879d41906dee006d606f66aa50e5e6653d12e5ec6a9f031b689c08ded5b9026
|
| 3 |
+
size 10907
|
baseline_code/EX1/doitgen.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:496fc5ac20435ea9cd43ccb5fd6b17069f5c050dc006ef30b5b6b6269045c209
|
| 3 |
+
size 7876
|
baseline_code/EX1/durbin.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:f851933a2a08e2f608f315c4fb7ef97f669ed780a727f988c3efc7da77fba07a
|
| 3 |
+
size 4556
|
baseline_code/EX1/fdtd-2d.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:122776f898b1dc4985490484a5bff705b5c4ff62559227bd321825062dc70564
|
| 3 |
+
size 4888
|
baseline_code/EX1/fft-strided-mach.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:695cbf053eaad0230777da8762b40548266a28e54847f4aff39cd75d33618025
|
| 3 |
+
size 6027
|
baseline_code/EX1/fft-transpose-mach.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:74c642a349effe452de559196bef8efd8567c62cf4646e5fa31d95d954d49313
|
| 3 |
+
size 11133
|
baseline_code/EX1/floyd-warshall.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d8308d403edddca72b78ba5fc2b73f7ec312a046eb5ceccbf99b25aa1dd22750
|
| 3 |
+
size 3652
|
baseline_code/EX1/gemm-blocked-mach.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ebe403bbc8c52fc1d9f967e306b3d9643a206468375bbf3988a8e8f6de3d758f
|
| 3 |
+
size 7872
|
baseline_code/EX1/gemm-ncubed-mach.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:353ce4f333ee23ded936e4a73b16669767538a1883dc2c04485e4d5ff26cffee
|
| 3 |
+
size 7968
|
baseline_code/EX1/gemm.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b14d84a4db438f2d5f14310d5b9cc504f6a27ad48957e2adcf58db2493ec23af
|
| 3 |
+
size 4613
|
baseline_code/EX1/gemver.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e5a4ff843064eac2357539e7a419421a6762d97c79ff5a3c337144357135d250
|
| 3 |
+
size 4768
|
baseline_code/EX1/gesummv.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d5e6483bb4f6191676421e1e818fdf44ae876cde803d909ac6f8f6a84acd84ad
|
| 3 |
+
size 4548
|
baseline_code/EX1/gramschmidt.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:1a4005f077a04f697f6040b3d037a8c936b3a740da5ce45fb8a38930f107ad09
|
| 3 |
+
size 4696
|
baseline_code/EX1/haccmk.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e7a54f14b270df58a512fc649495512802b11ebf823276d5c625a7d203a4917e
|
| 3 |
+
size 9296
|
baseline_code/EX1/heartwall.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:69c47fe806b06566d7f87965c7b286df61c16699a2ceade0f7636f1e7993af7b
|
| 3 |
+
size 31094
|
baseline_code/EX1/heat-3d.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3572c292253f2112ba0c20d25d918aecf754011c47760dd6c31378934fd39f67
|
| 3 |
+
size 4722
|
baseline_code/EX1/histo.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:a7341b8ad3417fc975416e73f1b78139d1a9b1d8965ca2612157f3f74ad03d62
|
| 3 |
+
size 15570
|
baseline_code/EX1/hotspot.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6d78620318a34fc4bc2f1ba534d44479280cfa52b757732ad95cefceae1e9765
|
| 3 |
+
size 4505
|
baseline_code/EX1/hotspot3D.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:21a78bb9094e5ebf0d126e633936c92f496c2370d9c320456f7431ed220531ac
|
| 3 |
+
size 3640
|
baseline_code/EX1/jacobi-1d.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6376d72e05642366675f2126de3ef05f3900a58d1598059efe2f7f16c321ec30
|
| 3 |
+
size 4548
|
baseline_code/EX1/jacobi-2d.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:8cc0d1dd38b8e4a0a45ae99c7ec66f11d74e04f403d00cac99ba368ba891d1b9
|
| 3 |
+
size 4531
|
baseline_code/EX1/kmeans.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b191b32f2c23d94e022f919ad8c6e58f909a5793de8ddb2dc5325e1a6469e108
|
| 3 |
+
size 30072
|
baseline_code/EX1/kmp-mach.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:b18dcb68c39d2dbf4ba9f1791d1a2b4e8c92a6ebe450a976e333a900dce1b76d
|
| 3 |
+
size 23383
|
baseline_code/EX1/lavaMD.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3addfde137e0a03f1dd6f243e604c3a65f5d2e8edac215774da8a4da496cf497
|
| 3 |
+
size 10624
|
baseline_code/EX1/lbm.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:afaadc92a8ecdd465f5d0bca99dac5c46edf4eafceb148380d133e76c83f17e9
|
| 3 |
+
size 21429
|
baseline_code/EX1/leukocyte.zip
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2c624ee4beec1b43341cbe78c3e98a5fa3683366b051f7dc43ec2870a329120c
|
| 3 |
+
size 26789
|