{
 "machine": {
  "safec": "safec 1.0.0",
  "clang": "Ubuntu clang version 21.1.8 (++20251221032922+2078da43e25a-1~exp1~20251221153059.70)",
  "rustc": "rustc 1.99.1 (58cf8f9c3 2026-10-13)",
  "go": "go version go1.27.3 linux/amd64",
  "zig": "0.17.0",
  "python": "Python 3.12.3",
  "torch": "2.10.0a0+b558c986e8.nv25.11",
  "tensorflow": "2.21.0",
  "numpy": "1.26.4",
  "cpu": "INTEL(R) XEON(R) PLATINUM 8570",
  "os": "Ubuntu 24.04.3 LTS",
  "cores": 12,
  "cuda": "Build cuda_13.0.r13.0/compiler.36424714_0",
  "gpu": "NVIDIA H200, 580.126.20"
 },
 "ml": {
  "small/safec_cpu": {
   ":train_ms": 24.336,
   ":last_loss": 100.333267,
   ":inference_ms_per_1000": 54.046
  },
  "small/safec_blas": {
   ":train_ms": 21.565,
   ":last_loss": 100.333267,
   ":inference_ms_per_1000": 55.292
  },
  "big/safec_cpu": {
   ":train_ms": 115.608,
   ":last_loss": 95.426926,
   ":inference_ms_per_200": 117.33
  },
  "big/safec_blas": {
   ":train_ms": 195.304,
   ":last_loss": 95.424927,
   ":inference_ms_per_200": 174.633
  },
  "small/safec_cpu_mkl": {
   ":train_ms": 25.862,
   ":last_loss": 100.333267,
   ":inference_ms_per_1000": 39.252
  },
  "big/safec_cpu_mkl": {
   ":train_ms": 101.323,
   ":last_loss": 95.427391,
   ":inference_ms_per_200": 61.273
  },
  "cuda/safec": {
   "small 128-256-64 plain": {
    "train_ms": 6.77,
    "last_loss": 100.333282,
    "throughput_samples_per_sec": 940164.82,
    "inference_ms_per_1000": 38.92,
    "checksum": -98.959379
   },
   "big 512-1024-256 plain": {
    "train_ms": 4.098,
    "last_loss": 95.427292,
    "throughput_samples_per_sec": 1559728.21,
    "inference_ms_per_200": 8.233,
    "checksum": 0.721703
   },
   "small 128-256-64 fused": {
    "train_ms": 6.154,
    "last_loss": 100.333282,
    "throughput_samples_per_sec": 1035796.32,
    "inference_ms_per_1000": 34.923,
    "checksum": -98.959379
   },
   "big 512-1024-256 fused": {
    "train_ms": 3.768,
    "last_loss": 95.427292,
    "throughput_samples_per_sec": 1685656.62,
    "inference_ms_per_200": 7.569,
    "checksum": 0.721703
   },
   "small 128-256-64 graph": {
    "train_ms": 3.812,
    "last_loss": 100.333282,
    "throughput_samples_per_sec": 1672568.99,
    "inference_ms_per_1000": 19.881,
    "checksum": -98.959379
   },
   "big 512-1024-256 graph": {
    "train_ms": 3.129,
    "last_loss": 95.427277,
    "throughput_samples_per_sec": 2045006.8,
    "inference_ms_per_200": 5.604,
    "checksum": 0.721703
   },
   "small 128-256-64 tf32": {
    "train_ms": 3.799,
    "last_loss": 100.334084,
    "throughput_samples_per_sec": 1684872.04,
    "inference_ms_per_1000": 19.61,
    "checksum": -99.043354
   },
   "big 512-1024-256 tf32": {
    "train_ms": 2.31,
    "last_loss": 90.685493,
    "throughput_samples_per_sec": 2770803.92,
    "inference_ms_per_200": 5.575,
    "checksum": 0.770854
   },
   "small 128-256-64 fp16": {
    "train_ms": 2.983,
    "last_loss": 100.335869,
    "throughput_samples_per_sec": 2145332.91,
    "inference_ms_per_1000": 19.653,
    "checksum": -98.999023
   },
   "big 512-1024-256 fp16": {
    "train_ms": 1.895,
    "last_loss": 93.444328,
    "throughput_samples_per_sec": 3366778.51,
    "inference_ms_per_200": 3.995,
    "checksum": 0.961304
   },
   "small 128-256-64 bf16": {
    "train_ms": 2.985,
    "last_loss": 99.95108,
    "throughput_samples_per_sec": 2143859.67,
    "inference_ms_per_1000": 19.664,
    "checksum": -89.84375
   },
   "big 512-1024-256 bf16": {
    "train_ms": 1.899,
    "last_loss": 62.00618,
    "throughput_samples_per_sec": 3365121.5,
    "inference_ms_per_200": 3.993,
    "checksum": -2.807617
   }
  },
  "small/pytorch_cpu": {
   "cpu:train_ms": 23.216,
   ":last_loss": 100.018272,
   "cpu:inference_ms_per_1000": 42.989
  },
  "small/pytorch_cuda": {
   "cuda:train_ms": 47.572,
   ":last_loss": 103.07663,
   "cuda:inference_ms_per_1000": 55.685
  },
  "small/pytorch_cuda_fp16": {
   "cuda:train_ms": 49.1,
   ":last_loss": 103.0625,
   "cuda:inference_ms_per_1000": 52.521
  },
  "small/pytorch_cuda_bf16": {
   "cuda:train_ms": 48.143,
   ":last_loss": 103.0,
   "cuda:inference_ms_per_1000": 52.353
  },
  "big/pytorch_cpu": {
   "cpu:train_ms": 63.148,
   ":last_loss": 106.544487,
   "cpu:inference_ms_per_200": 53.564
  },
  "big/pytorch_cuda": {
   "cuda:train_ms": 18.45,
   ":last_loss": 871.826538,
   "cuda:inference_ms_per_200": 11.811
  },
  "big/pytorch_cuda_fp16": {
   "cuda:train_ms": 17.259,
   ":last_loss": 1281.0,
   "cuda:inference_ms_per_200": 10.459
  },
  "big/pytorch_cuda_bf16": {
   "cuda:train_ms": 16.742,
   ":last_loss": 91.5,
   "cuda:inference_ms_per_200": 10.52
  },
  "small/tensorflow_cpu": {
   "cpu:train_ms": 122.926,
   ":last_loss": 106.182999,
   "cpu:inference_ms_per_1000": 264.942
  },
  "big/tensorflow_cpu": {
   "cpu:train_ms": 228.456,
   ":last_loss": 316.611603,
   "cpu:inference_ms_per_200": 236.028
  },
  "small/pytorch_cuda_fp32": {
   "cuda:train_ms": 51.266,
   ":last_loss": 103.071129,
   "cuda:inference_ms_per_1000": 56.562,
   "note": "TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=0 (the NGC container defaults to 1, i.e. TF32), best of 3"
  },
  "big/pytorch_cuda_fp32": {
   "cuda:train_ms": 15.9,
   ":last_loss": 1754.826172,
   "cuda:inference_ms_per_200": 11.78,
   "note": "TORCH_ALLOW_TF32_CUBLAS_OVERRIDE=0, best of 3"
  }
 },
 "fib": {
  "safec": {
   "debug": 0.2082,
   "release": 0.0892,
   "rss_mb": 8.8,
   "compile_s": 0.119
  },
  "c": {
   "debug": 0.1849,
   "release": 0.0869,
   "rss_mb": 8.8,
   "compile_s": 0.099
  },
  "cpp": {
   "debug": 0.1851,
   "release": 0.0886,
   "rss_mb": 10.5,
   "compile_s": 0.127
  },
  "rust": {
   "debug": 0.263,
   "release": 0.0924,
   "rss_mb": 10.5,
   "compile_s": 0.162
  },
  "zig": {
   "debug": 0.2775,
   "release": 0.072,
   "rss_mb": 2.8,
   "compile_s": 14.478
  },
  "go": {
   "debug": 0.1629,
   "release": 0.1324,
   "rss_mb": 2.8,
   "compile_s": 0.064
  },
  "python": {
   "debug": 2.4344,
   "release": 2.4316,
   "rss_mb": 19.2,
   "compile_s": null
  }
 },
 "nbody": {
  "safec": {
   "debug": 0.5092,
   "release": 0.1446,
   "rss_mb": 8.8,
   "compile_s": 0.143
  },
  "c": {
   "debug": 0.3223,
   "release": 0.1459,
   "rss_mb": 8.8,
   "compile_s": 0.133
  },
  "cpp": {
   "debug": 0.3212,
   "release": 0.1466,
   "rss_mb": 10.5,
   "compile_s": 0.26
  },
  "rust": {
   "debug": 0.5972,
   "release": 0.1547,
   "rss_mb": 10.5,
   "compile_s": 0.181
  },
  "zig": {
   "debug": 0.3421,
   "release": 0.1541,
   "rss_mb": 2.8,
   "compile_s": 14.796
  },
  "go": {
   "debug": 0.3515,
   "release": 0.1898,
   "rss_mb": 2.8,
   "compile_s": 0.059
  },
  "python": {
   "debug": 9.3038,
   "release": 9.3366,
   "rss_mb": 19.2,
   "compile_s": null
  }
 },
 "binarytrees": {
  "safec": {
   "debug": 0.5135,
   "release": 0.2816,
   "rss_mb": 33.2,
   "compile_s": 0.14
  },
  "c": {
   "debug": 1.313,
   "release": 1.1751,
   "rss_mb": 42.0,
   "compile_s": 0.11
  },
  "cpp": {
   "debug": 1.693,
   "release": 1.3632,
   "rss_mb": 43.8,
   "compile_s": 0.136
  },
  "rust": {
   "debug": 3.0609,
   "release": 1.4048,
   "rss_mb": 43.8,
   "compile_s": 0.171
  },
  "zig": {
   "debug": 6.4545,
   "release": 1.2527,
   "rss_mb": 42.0,
   "compile_s": 13.758
  },
  "go": {
   "debug": 6.8466,
   "release": 3.5104,
   "rss_mb": 34.8,
   "compile_s": 0.043
  },
  "python": {
   "debug": 29.4874,
   "release": 29.1657,
   "rss_mb": 92.0,
   "compile_s": null
  },
  "safec_heap": {
   "debug": 1.3327,
   "release": 0.9452,
   "rss_mb": 47.1,
   "compile_s": 0.135
  }
 },
 "binarytrees_mt": {
  "safec": {
   "release": 0.0921,
   "rss_mb": 95.9,
   "compile_s": 0.239
  },
  "c": {
   "release": 0.4175,
   "rss_mb": 161.0,
   "compile_s": 0.116
  },
  "cpp": {
   "release": 0.4635,
   "rss_mb": 168.0,
   "compile_s": 0.371
  },
  "rust": {
   "release": 0.457,
   "rss_mb": 164.5,
   "compile_s": 0.331
  },
  "zig": {
   "error": "command failed (1): /home/work/.local/zig-x86_64-linux-0.17.0/zig build-exe -O ReleaseFast /home/work/safec-docs/public/benchmarks/binarytrees_mt/zig/binarytrees_mt.zig -femit-bin=/home/work/safec-doc"
  },
  "go": {
   "release": 0.4097,
   "rss_mb": 139.9,
   "compile_s": 0.051
  },
  "python": {
   "release": 76.8511,
   "rss_mb": 320.8,
   "compile_s": null
  },
  "safec_heap": {
   "release": 0.333,
   "rss_mb": 202.9,
   "compile_s": 0.159
  }
 },
 "simd": {
  "safec": {
   "scalar": 0.1213,
   "vec": 0.1108
  },
  "c": {
   "scalar": 0.115,
   "vec": 0.1181
  },
  "cpp": {
   "scalar": 0.1246,
   "vec": 0.1165
  },
  "rust": {
   "scalar": 0.1155,
   "vec": 0.1122
  },
  "zig": {
   "scalar": 0.1041,
   "vec": 0.0976
  },
  "go": {
   "scalar": 0.1185
  },
  "python": {
   "scalar": 4.0799,
   "vec": 0.3329
  }
 },
 "collections": {
  "safec": {
   "vec push": 3.333,
   "vec sort": 76.3,
   "map insert": 56.294,
   "map get": 20.093,
   "list push+pop": 13.282,
   "queue enq+deq": 3.296,
   "bst insert": 348.75,
   "bst get": 350.97
  },
  "c": {
   "vec push": 1.945,
   "vec sort": 149.042,
   "map insert": 53.944,
   "map get": 18.611,
   "list push+pop": 27.242,
   "queue enq+deq": 6.641,
   "bst insert": 312.025,
   "bst get": 337.789
  }
 },
 "memalloc": {
  "std_ns_per_pair": 10.4,
  "malloc_ns_per_pair": 8.83,
  "std_ms": 119.67,
  "malloc_ms": 126.05
 },
 "web": {
  "safec": {
   "rps": 33730.4,
   "p50_ms": 1,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 14.2
  },
  "c": {
   "rps": 32196.22,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 14.0
  },
  "cpp": {
   "rps": 32522.7,
   "p50_ms": 1,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 15.7
  },
  "zig": {
   "error": "command failed (1): /home/work/.local/zig-x86_64-linux-0.17.0/zig build-exe -O ReleaseFast /home/work/safec-docs/public/benchmarks/webservice/zig/server.zig -femit-bin=/home/work/safec-docs/public/ben"
  },
  "go": {
   "rps": 32218.62,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 25.3
  },
  "rust": {
   "rps": 32592.03,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 13.6
  },
  "python": {
   "rps": 2887.48,
   "p50_ms": 17,
   "p99_ms": 30,
   "failed": 0,
   "rss_mb": 60.8
  }
 },
 "conv": {
  "safec_cpu": {
   "conv64 b=1 cpu": {
    "ms_per_image": 0.854,
    "checksum": 528.229
   },
   "unet128 b=1 cpu": {
    "ms_per_image": 3.315,
    "checksum": 994785.1153
   }
  },
  "safec_cpu_mkl": {
   "conv64 b=1 cpu": {
    "ms_per_image": 0.71,
    "checksum": 528.229
   },
   "unet128 b=1 cpu": {
    "ms_per_image": 2.65,
    "checksum": 994785.1082
   }
  },
  "safec_cuda": {
   "conv64 b=1 fp32": {
    "ms_per_batch": 0.0181,
    "ms_per_image": 0.01808
   },
   "conv64 b=32 fp32": {
    "ms_per_batch": 0.1642,
    "ms_per_image": 0.00513
   },
   "unet128 b=1 fp32": {
    "ms_per_batch": 0.1557,
    "ms_per_image": 0.15567
   },
   "unet128 b=1 graph fp32": {
    "ms_per_batch": 0.1437,
    "ms_per_image": 0.14372
   },
   "unet128 b=16 fp32": {
    "ms_per_batch": 0.5342,
    "ms_per_image": 0.03339
   },
   "conv64 b=1 tf32": {
    "ms_per_batch": 0.0181,
    "ms_per_image": 0.01814
   },
   "conv64 b=32 tf32": {
    "ms_per_batch": 0.0838,
    "ms_per_image": 0.00262
   },
   "unet128 b=1 tf32": {
    "ms_per_batch": 0.104,
    "ms_per_image": 0.10404
   },
   "unet128 b=1 graph tf32": {
    "ms_per_batch": 0.0874,
    "ms_per_image": 0.08743
   },
   "unet128 b=16 tf32": {
    "ms_per_batch": 0.4537,
    "ms_per_image": 0.02836
   }
  },
  "pytorch_cpu": {
   "conv64 b=1 cpu": {
    "ms_per_batch": 0.4239,
    "ms_per_image": 0.42394
   },
   "unet128 b=1 cpu": {
    "ms_per_batch": 3.2874,
    "ms_per_image": 3.28736
   }
  },
  "pytorch_cuda": {
   "conv64 b=1 fp32": {
    "ms_per_batch": 0.0225,
    "ms_per_image": 0.02249
   },
   "conv64 b=32 fp32": {
    "ms_per_batch": 0.257,
    "ms_per_image": 0.00803
   },
   "unet128 b=1 fp32": {
    "ms_per_batch": 0.2082,
    "ms_per_image": 0.20823
   },
   "unet128 b=16 fp32": {
    "ms_per_batch": 0.8766,
    "ms_per_image": 0.05479
   },
   "conv64 b=1 tf32": {
    "ms_per_batch": 0.038,
    "ms_per_image": 0.03799
   },
   "conv64 b=32 tf32": {
    "ms_per_batch": 0.0976,
    "ms_per_image": 0.00305
   },
   "unet128 b=1 tf32": {
    "ms_per_batch": 0.2689,
    "ms_per_image": 0.26892
   },
   "unet128 b=16 tf32": {
    "ms_per_batch": 0.5566,
    "ms_per_image": 0.03479
   },
   "conv64 b=1 fp32 heuristic": {
    "ms_per_batch": 0.0223,
    "ms_per_image": 0.02232
   },
   "conv64 b=32 fp32 heuristic": {
    "ms_per_batch": 0.257,
    "ms_per_image": 0.00803
   },
   "unet128 b=1 fp32 heuristic": {
    "ms_per_batch": 0.2074,
    "ms_per_image": 0.2074
   },
   "unet128 b=16 fp32 heuristic": {
    "ms_per_batch": 0.8763,
    "ms_per_image": 0.05477
   }
  }
 },
 "fft": {
  "safec_cpu": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 1.205
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 1.081
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 2.963
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 2.313
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 3.063
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 2.012
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 3.163
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 2.358
   }
  },
  "safec_cpu_mkl": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 1.179
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 0.837
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 1.647
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 1.364
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 2.85
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 1.112
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 1.602
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 0.812
   }
  },
  "safec_cuda": {
   "fft1d n=1048576 c128 cuda": {
    "ms_per_call": 1.2539
   },
   "fft2d 1024x1024 c128 cuda": {
    "ms_per_call": 1.2572
   },
   "batch 1024x4096 c2c cuda": {
    "ms_per_call": 0.0244
   },
   "rfft 1024x4096 r2c cuda": {
    "ms_per_call": 0.0109
   }
  },
  "pytorch_cpu": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 1.1811
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 0.654
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 3.1884
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 0.2876
   }
  },
  "pytorch_cuda": {
   "fft1d n=1048576 c128 cuda": {
    "ms_per_call": 2.0526
   },
   "fft2d 1024x1024 c128 cuda": {
    "ms_per_call": 2.0447
   },
   "batch 1024x4096 c2c cuda": {
    "ms_per_call": 0.0238
   },
   "rfft 1024x4096 r2c cuda": {
    "ms_per_call": 0.012
   }
  },
  "fftw_1t": {
   "fft1d n=1048576 c128 st": {
    "ms_per_call": 11.394
   },
   "fft2d 1024x1024 c128 st": {
    "ms_per_call": 7.273
   },
   "batch 1024x4096 c2c st": {
    "ms_per_call": 10.667
   },
   "rfft 1024x4096 r2c st": {
    "ms_per_call": 7.351
   },
   "fft1d n=1048576 c64 st": {
    "ms_per_call": 8.719
   },
   "fft2d 1024x1024 c64 st": {
    "ms_per_call": 4.292
   },
   "batch 1024x4096 c2c f32 st": {
    "ms_per_call": 6.194
   },
   "rfft 1024x4096 r2c f32 st": {
    "ms_per_call": 6.5
   }
  },
  "fftw_16t": {
   "fft1d n=1048576 c128 mt": {
    "ms_per_call": 2.354
   },
   "fft2d 1024x1024 c128 mt": {
    "ms_per_call": 2.588
   },
   "batch 1024x4096 c2c mt": {
    "ms_per_call": 2.947
   },
   "rfft 1024x4096 r2c mt": {
    "ms_per_call": 2.017
   },
   "fft1d n=1048576 c64 mt": {
    "ms_per_call": 2.456
   },
   "fft2d 1024x1024 c64 mt": {
    "ms_per_call": 1.862
   },
   "batch 1024x4096 c2c f32 mt": {
    "ms_per_call": 1.755
   },
   "rfft 1024x4096 r2c f32 mt": {
    "ms_per_call": 2.079
   }
  },
  "kfr": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 6.35
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 5.147
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 12.34
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 7.324
   },
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 42.831
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 6.583
   },
   "biquad n=1048576": {
    "ms_per_call": 4.45
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 88.977
   },
   "resample draft n=1048576 44100->48000": {
    "ms_per_call": 58.096
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 3.311
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 3.068
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 6.03
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 3.735
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 37.111
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 3.278
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 4.266
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 53.364
   },
   "resample draft n=1048576 44100->48000 f32": {
    "ms_per_call": 60.623
   }
  },
  "rust": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 8.781
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 14.826
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 4.119
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 7.754
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 657.65
   }
  }
 },
 "dsp": {
  "safec_cpu": {
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 7.459,
    "checksum": 965.503541
   },
   "fir block n=1048576 taps=127": {
    "ms_per_call": 0.73,
    "checksum": 965.503541
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 4.536,
    "checksum": 11438.877046
   },
   "biquad n=1048576": {
    "ms_per_call": 0.782,
    "checksum": -646.722217
   },
   "stft n=1048576 frame=1024 hop=256": {
    "ms_per_call": 5.691,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 7.38,
    "checksum": -704.289117
   },
   "conv2d 1024x1024 k=5x5": {
    "ms_per_call": 0.601,
    "checksum": 205.652439
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 3.826,
    "checksum": 965.503677
   },
   "fir block n=1048576 taps=127 f32": {
    "ms_per_call": 0.649,
    "checksum": 965.503676
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 4.233,
    "checksum": 11438.881763
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 0.773,
    "checksum": -646.722416
   },
   "stft n=1048576 frame=1024 hop=256 f32": {
    "ms_per_call": 4.883,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 11.129,
    "checksum": -704.28909
   }
  },
  "safec_cpu_mkl": {
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 7.374,
    "checksum": 965.503541
   },
   "fir block n=1048576 taps=127": {
    "ms_per_call": 0.686,
    "checksum": 965.503541
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 3.817,
    "checksum": 11438.877046
   },
   "biquad n=1048576": {
    "ms_per_call": 0.782,
    "checksum": -646.722217
   },
   "stft n=1048576 frame=1024 hop=256": {
    "ms_per_call": 5.471,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 8.029,
    "checksum": -704.289117
   },
   "conv2d 1024x1024 k=5x5": {
    "ms_per_call": 0.715,
    "checksum": 205.652439
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 3.855,
    "checksum": 965.503677
   },
   "fir block n=1048576 taps=127 f32": {
    "ms_per_call": 0.673,
    "checksum": 965.503676
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 2.205,
    "checksum": 11438.877545
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 0.785,
    "checksum": -646.722416
   },
   "stft n=1048576 frame=1024 hop=256 f32": {
    "ms_per_call": 4.145,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 10.764,
    "checksum": -704.28909
   }
  },
  "scipy": {
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 27.497
   },
   "fir block n=1048576 taps=127": {
    "ms_per_call": 25.093
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 60.509
   },
   "biquad n=1048576": {
    "ms_per_call": 6.989
   },
   "stft n=1048576 frame=1024 hop=256": {
    "ms_per_call": 92.369
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 17.51
   },
   "conv2d 1024x1024 k=5x5": {
    "ms_per_call": 53.208
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 16.04
   },
   "fir block n=1048576 taps=127 f32": {
    "ms_per_call": 15.695
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 32.152
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 6.342
   },
   "stft n=1048576 frame=1024 hop=256 f32": {
    "ms_per_call": 95.062
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 16.534
   }
  },
  "kfr": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 6.355
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 5.149
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 12.082
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 7.391
   },
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 42.818
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 6.57
   },
   "biquad n=1048576": {
    "ms_per_call": 4.446
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 88.398
   },
   "resample draft n=1048576 44100->48000": {
    "ms_per_call": 58.135
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 3.289
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 3.071
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 6.005
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 3.766
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 37.163
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 3.28
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 4.265
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 53.183
   },
   "resample draft n=1048576 44100->48000 f32": {
    "ms_per_call": 60.616
   }
  },
  "q_madronalib": {
   "biquad q n=1048576 f32": {
    "ms_per_call": 2.814,
    "checksum": 996.364585
   },
   "biquad madronalib n=1048576 f32": {
    "ms_per_call": 5.367,
    "checksum": 996.365938
   }
  },
  "rust": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 8.88
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 14.922
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 4.169
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 7.187
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 656.88
   }
  }
 },
 "attention": {
  "safec_cuda": {
   "flash fp16 d64 s1024 c0": {
    "ms_per_call": 0.0432,
    "tflops": 397.96
   },
   "flash fp16 d64 s1024 c1": {
    "ms_per_call": 0.0359,
    "tflops": 238.91
   },
   "flash fp16 d64 s4096 c0": {
    "ms_per_call": 0.5719,
    "tflops": 480.65
   },
   "flash fp16 d64 s4096 c1": {
    "ms_per_call": 0.3363,
    "tflops": 407.45
   },
   "flash fp16 d128 s1024 c0": {
    "ms_per_call": 0.0601,
    "tflops": 570.69
   },
   "flash fp16 d128 s1024 c1": {
    "ms_per_call": 0.0487,
    "tflops": 352.2
   },
   "flash fp16 d128 s4096 c0": {
    "ms_per_call": 0.8314,
    "tflops": 660.91
   },
   "flash fp16 d128 s4096 c1": {
    "ms_per_call": 0.4803,
    "tflops": 572.09
   },
   "flash bf16 d64 s1024 c0": {
    "ms_per_call": 0.0429,
    "tflops": 400.46
   },
   "flash bf16 d64 s1024 c1": {
    "ms_per_call": 0.0359,
    "tflops": 238.97
   },
   "flash bf16 d64 s4096 c0": {
    "ms_per_call": 0.5688,
    "tflops": 483.27
   },
   "flash bf16 d64 s4096 c1": {
    "ms_per_call": 0.3361,
    "tflops": 408.93
   },
   "flash bf16 d128 s1024 c0": {
    "ms_per_call": 0.0585,
    "tflops": 586.27
   },
   "flash bf16 d128 s1024 c1": {
    "ms_per_call": 0.0483,
    "tflops": 353.73
   },
   "flash bf16 d128 s4096 c0": {
    "ms_per_call": 0.7982,
    "tflops": 688.04
   },
   "flash bf16 d128 s4096 c1": {
    "ms_per_call": 0.4663,
    "tflops": 588.62
   },
   "flash fp32 d64 s1024 c0": {
    "ms_per_call": 0.3867,
    "tflops": 44.42
   },
   "flash fp32 d64 s1024 c1": {
    "ms_per_call": 0.2632,
    "tflops": 32.63
   },
   "flash fp32 d64 s4096 c0": {
    "ms_per_call": 5.8086,
    "tflops": 47.3
   },
   "flash fp32 d64 s4096 c1": {
    "ms_per_call": 3.1128,
    "tflops": 44.15
   },
   "flash fp32 d128 s1024 c0": {
    "ms_per_call": 0.7481,
    "tflops": 45.93
   },
   "flash fp32 d128 s1024 c1": {
    "ms_per_call": 0.502,
    "tflops": 33.91
   },
   "flash fp32 d128 s4096 c0": {
    "ms_per_call": 11.3259,
    "tflops": 48.53
   },
   "flash fp32 d128 s4096 c1": {
    "ms_per_call": 6.0671,
    "tflops": 45.02
   },
   "flash bf16 gqa32/8 d128 s4096 c1": {
    "ms_per_call": 0.8926,
    "tflops": 614.68
   },
   "linear fp32 d64 s4096 c0": {
    "ms_per_call": 0.2307
   },
   "linear fp32 d64 s4096 c1": {
    "ms_per_call": 0.3847
   },
   "linear fp32 d64 s16384 c0": {
    "ms_per_call": 0.8681
   },
   "linear fp32 d64 s16384 c1": {
    "ms_per_call": 1.4385
   },
   "linear fp32 d128 s4096 c0": {
    "ms_per_call": 0.6329
   },
   "linear fp32 d128 s4096 c1": {
    "ms_per_call": 1.0303
   },
   "linear bf16 d64 s4096 c0": {
    "ms_per_call": 0.2284
   },
   "linear bf16 d64 s4096 c1": {
    "ms_per_call": 0.3924
   },
   "linear bf16 d64 s16384 c0": {
    "ms_per_call": 0.8526
   },
   "linear bf16 d64 s16384 c1": {
    "ms_per_call": 1.4759
   },
   "linear bf16 d128 s4096 c0": {
    "ms_per_call": 0.6188
   },
   "linear bf16 d128 s4096 c1": {
    "ms_per_call": 1.0093
   }
  },
  "pytorch_cuda": {
   "flash fp16 d64 s1024 c0": {
    "ms_per_call": 0.0437,
    "tflops": 392.99
   },
   "flash fp16 d64 s1024 c1": {
    "ms_per_call": 0.0352,
    "tflops": 244.2
   },
   "flash fp16 d64 s4096 c0": {
    "ms_per_call": 0.575,
    "tflops": 478.02
   },
   "flash fp16 d64 s4096 c1": {
    "ms_per_call": 0.332,
    "tflops": 413.98
   },
   "flash fp16 d128 s1024 c0": {
    "ms_per_call": 0.0599,
    "tflops": 573.53
   },
   "flash fp16 d128 s1024 c1": {
    "ms_per_call": 0.0498,
    "tflops": 344.92
   },
   "flash fp16 d128 s4096 c0": {
    "ms_per_call": 0.8215,
    "tflops": 669.18
   },
   "flash fp16 d128 s4096 c1": {
    "ms_per_call": 0.4781,
    "tflops": 574.91
   },
   "flash bf16 d64 s1024 c0": {
    "ms_per_call": 0.0433,
    "tflops": 396.13
   },
   "flash bf16 d64 s1024 c1": {
    "ms_per_call": 0.035,
    "tflops": 245.27
   },
   "flash bf16 d64 s4096 c0": {
    "ms_per_call": 0.573,
    "tflops": 479.4
   },
   "flash bf16 d64 s4096 c1": {
    "ms_per_call": 0.3309,
    "tflops": 414.0
   },
   "flash bf16 d128 s1024 c0": {
    "ms_per_call": 0.0587,
    "tflops": 584.86
   },
   "flash bf16 d128 s1024 c1": {
    "ms_per_call": 0.0494,
    "tflops": 347.11
   },
   "flash bf16 d128 s4096 c0": {
    "ms_per_call": 0.786,
    "tflops": 698.93
   },
   "flash bf16 d128 s4096 c1": {
    "ms_per_call": 0.4634,
    "tflops": 590.8
   },
   "flash fp32 d64 s1024 c0": {
    "ms_per_call": 0.5192,
    "tflops": 32.82
   },
   "flash fp32 d64 s1024 c1": {
    "ms_per_call": 0.3532,
    "tflops": 24.32
   },
   "flash fp32 d64 s4096 c0": {
    "ms_per_call": 7.7843,
    "tflops": 35.31
   },
   "flash fp32 d64 s4096 c1": {
    "ms_per_call": 4.1783,
    "tflops": 32.83
   },
   "flash fp32 d128 s1024 c0": {
    "ms_per_call": 0.7885,
    "tflops": 43.58
   },
   "flash fp32 d128 s1024 c1": {
    "ms_per_call": 0.4848,
    "tflops": 35.44
   },
   "flash fp32 d128 s4096 c0": {
    "ms_per_call": 11.8905,
    "tflops": 46.23
   },
   "flash fp32 d128 s4096 c1": {
    "ms_per_call": 6.1937,
    "tflops": 44.38
   },
   "flash bf16 gqa32/8 d128 s4096 c1": {
    "ms_per_call": 0.8845,
    "tflops": 621.55
   },
   "linear fp32 d64 s4096 c0": {
    "ms_per_call": 0.3497
   },
   "linear fp32 ieee d64 s4096 c0": {
    "ms_per_call": 0.5004
   },
   "linear fp32 d64 s4096 c1": {
    "ms_per_call": 0.8793
   },
   "linear fp32 ieee d64 s4096 c1": {
    "ms_per_call": 0.9975
   },
   "linear fla fp32 d64 s4096 c1": {
    "ms_per_call": 0.815,
    "maxdiff": 0.00121
   },
   "linear fp32 d64 s16384 c0": {
    "ms_per_call": 1.4064
   },
   "linear fp32 ieee d64 s16384 c0": {
    "ms_per_call": 1.9723
   },
   "linear fp32 d64 s16384 c1": {
    "ms_per_call": 3.3205
   },
   "linear fp32 ieee d64 s16384 c1": {
    "ms_per_call": 3.7411
   },
   "linear fla fp32 d64 s16384 c1": {
    "ms_per_call": 3.0786,
    "maxdiff": 0.0012
   },
   "linear fp32 d128 s4096 c0": {
    "ms_per_call": 0.6287
   },
   "linear fp32 ieee d128 s4096 c0": {
    "ms_per_call": 0.934
   },
   "linear fp32 d128 s4096 c1": {
    "ms_per_call": 1.6911
   },
   "linear fp32 ieee d128 s4096 c1": {
    "ms_per_call": 2.043
   },
   "linear fla fp32 d128 s4096 c1": {
    "ms_per_call": 1.5867,
    "maxdiff": 0.00121
   },
   "linear bf16 d64 s4096 c0": {
    "ms_per_call": 0.5038
   },
   "linear bf16 d64 s4096 c1": {
    "ms_per_call": 1.0312
   },
   "linear fla bf16 d64 s4096 c1": {
    "ms_per_call": 0.6434,
    "maxdiff": 0.00398
   },
   "linear bf16 d64 s16384 c0": {
    "ms_per_call": 1.9759
   },
   "linear bf16 d64 s16384 c1": {
    "ms_per_call": 3.891
   },
   "linear fla bf16 d64 s16384 c1": {
    "ms_per_call": 2.383,
    "maxdiff": 0.00393
   },
   "linear bf16 d128 s4096 c0": {
    "ms_per_call": 0.9215
   },
   "linear bf16 d128 s4096 c1": {
    "ms_per_call": 1.9817
   },
   "linear fla bf16 d128 s4096 c1": {
    "ms_per_call": 1.152,
    "maxdiff": 0.00392
   }
  }
 },
 "ml_nn": {
  "safec_cpu": {
   "mlp_small": {
    "train_ms": 6.0465,
    "last_loss": 0.0297225
   },
   "mlp_big": {
    "train_ms": 53.7706,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 24.1021,
    "last_loss": 31.5691
   },
   "gan": {
    "train_ms": 14.1811,
    "last_loss": 1.40552
   },
   "dqn": {
    "train_ms": 82.9076,
    "last_loss": 0.138698
   },
   "ppo": {
    "train_ms": 43.5158,
    "last_loss": 48.7648
   },
   "lora": {
    "train_ms": 19.7919,
    "last_loss": 0.0219908
   },
   "conv_ae": {
    "train_ms": 72.2017,
    "last_loss": 0.255294
   }
  },
  "safec_cpu_mkl": {
   "mlp_small": {
    "train_ms": 5.69948,
    "last_loss": 0.0297225
   },
   "mlp_big": {
    "train_ms": 39.7831,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 20.5492,
    "last_loss": 31.5691
   },
   "gan": {
    "train_ms": 12.0997,
    "last_loss": 1.40552
   },
   "dqn": {
    "train_ms": 67.8341,
    "last_loss": 0.138698
   },
   "ppo": {
    "train_ms": 35.3655,
    "last_loss": 48.7648
   },
   "lora": {
    "train_ms": 18.5055,
    "last_loss": 0.0219908
   },
   "conv_ae": {
    "train_ms": 65.3947,
    "last_loss": 0.255294
   }
  },
  "safec_cuda": {
   "mlp_small": {
    "train_ms": 82.7629,
    "last_loss": 0.0297225
   },
   "mlp_big": {
    "train_ms": 9.8884,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 65.5617,
    "last_loss": 31.5691
   },
   "gan": {
    "train_ms": 116.84,
    "last_loss": 1.40552
   },
   "dqn": {
    "train_ms": 402.048,
    "last_loss": 0.138697
   },
   "ppo": {
    "train_ms": 168.715,
    "last_loss": 48.7648
   },
   "lora": {
    "train_ms": 184.215,
    "last_loss": 0.0219907
   },
   "conv_ae": {
    "train_ms": 13.9268,
    "last_loss": 0.255294
   }
  },
  "pytorch_cpu": {
   "mlp_small": {
    "train_ms": 217.26,
    "last_loss": 0.0331768
   },
   "mlp_big": {
    "train_ms": 72.979,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 163.502,
    "last_loss": 50.6271
   },
   "gan": {
    "train_ms": 209.274,
    "last_loss": 1.42086
   },
   "dqn": {
    "train_ms": 1170.296,
    "last_loss": 0.126866
   },
   "ppo": {
    "train_ms": 453.073,
    "last_loss": 41.1927
   },
   "lora": {
    "train_ms": 303.985,
    "last_loss": 0.0405097
   },
   "conv_ae": {
    "train_ms": 170.561,
    "last_loss": 0.484414
   }
  },
  "pytorch_cuda": {
   "mlp_small": {
    "train_ms": 1086.569,
    "last_loss": 0.0556656
   },
   "mlp_big": {
    "train_ms": 110.98,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 774.459,
    "last_loss": 50.4145
   },
   "gan": {
    "train_ms": 1320.124,
    "last_loss": 1.47957
   },
   "dqn": {
    "train_ms": 2600.045,
    "last_loss": 0.223958
   },
   "ppo": {
    "train_ms": 2017.787,
    "last_loss": 36.4079
   },
   "lora": {
    "train_ms": 1804.658,
    "last_loss": 0.0498237
   },
   "conv_ae": {
    "train_ms": 212.94,
    "last_loss": 0.500654
   }
  }
 }
}