{
 "machine": {
  "safec": "safec 1.0.0",
  "clang": "clang version 21.1.8 (https://github.com/llvm/llvm-project 2078da43e25a4623cab2d0d60decddf709aaea28)",
  "rustc": "rustc 1.100.0-nightly (bba531001 2026-09-20)",
  "go": null,
  "zig": null,
  "python": "Python 3.12.3",
  "torch": "2.9.1+rocm6.4",
  "tensorflow": "2.21.0",
  "numpy": "2.5.4",
  "mlx": "Traceback (most recent call last):",
  "cpu": "AMD Ryzen 7 7800X3D 8-Core Processor",
  "os": "Ubuntu 24.04.5 LTS",
  "cores": 16,
  "gpu": "AMD Instinct MI100"
 },
 "fib": {
  "safec": {
   "debug": 0.1065,
   "release": 0.042,
   "rss_mb": 1.9,
   "compile_s": 0.036
  },
  "c": {
   "debug": 0.0978,
   "release": 0.0418,
   "rss_mb": 1.9,
   "compile_s": 0.03
  },
  "cpp": {
   "debug": 0.0975,
   "release": 0.0423,
   "rss_mb": 3.8,
   "compile_s": 0.049
  },
  "rust": {
   "debug": 0.1235,
   "release": 0.0417,
   "rss_mb": 1.9,
   "compile_s": 0.05
  },
  "zig": {
   "debug": 0.1947,
   "release": 0.0427,
   "rss_mb": 0.6,
   "compile_s": 8.175
  },
  "go": {
   "debug": 0.0939,
   "release": 0.0887,
   "rss_mb": 2.0,
   "compile_s": 0.025
  },
  "python": {
   "debug": 2.0073,
   "release": 2.0158,
   "rss_mb": 10.4,
   "compile_s": null
  }
 },
 "nbody": {
  "safec": {
   "debug": 0.4855,
   "release": 0.0981,
   "rss_mb": 2.0,
   "compile_s": 0.044
  },
  "c": {
   "debug": 0.2109,
   "release": 0.0988,
   "rss_mb": 2.0,
   "compile_s": 0.043
  },
  "cpp": {
   "debug": 0.2096,
   "release": 0.0987,
   "rss_mb": 3.8,
   "compile_s": 0.097
  },
  "rust": {
   "debug": 0.4492,
   "release": 0.1067,
   "rss_mb": 2.0,
   "compile_s": 0.072
  },
  "zig": {
   "debug": 0.2372,
   "release": 0.1095,
   "rss_mb": 0.5,
   "compile_s": 8.249
  },
  "go": {
   "debug": 0.2229,
   "release": 0.1056,
   "rss_mb": 2.0,
   "compile_s": 0.026
  },
  "python": {
   "debug": 7.7396,
   "release": 7.6301,
   "rss_mb": 10.9,
   "compile_s": null
  }
 },
 "binarytrees": {
  "safec": {
   "debug": 0.2625,
   "release": 0.1317,
   "rss_mb": 25.9,
   "compile_s": 0.044
  },
  "c": {
   "debug": 0.9714,
   "release": 0.8098,
   "rss_mb": 33.9,
   "compile_s": 0.039
  },
  "cpp": {
   "debug": 1.1837,
   "release": 0.9242,
   "rss_mb": 35.9,
   "compile_s": 0.054
  },
  "rust": {
   "debug": 1.7701,
   "release": 0.9137,
   "rss_mb": 34.0,
   "compile_s": 0.064
  },
  "zig": {
   "debug": 5.0526,
   "release": 0.8542,
   "rss_mb": 33.6,
   "compile_s": 8.219
  },
  "go": {
   "debug": 1.8291,
   "release": 0.7183,
   "rss_mb": 36.8,
   "compile_s": 0.035
  },
  "python": {
   "debug": 18.0346,
   "release": 18.0495,
   "rss_mb": 82.6,
   "compile_s": null
  },
  "safec_heap": {
   "debug": 0.8794,
   "release": 0.6162,
   "rss_mb": 38.8,
   "compile_s": 0.043
  }
 },
 "binarytrees_mt": {
  "safec": {
   "release": 0.0526,
   "rss_mb": 98.4,
   "compile_s": 0.074
  },
  "c": {
   "release": 0.2261,
   "rss_mb": 161.8,
   "compile_s": 0.041
  },
  "cpp": {
   "release": 0.2271,
   "rss_mb": 163.5,
   "compile_s": 0.141
  },
  "rust": {
   "release": 0.2324,
   "rss_mb": 161.9,
   "compile_s": 0.144
  },
  "zig": {
   "release": 0.2259,
   "rss_mb": 163.8,
   "compile_s": 8.253
  },
  "go": {
   "release": 0.2484,
   "rss_mb": 96.4,
   "compile_s": 0.035
  },
  "python": {
   "release": 22.9576,
   "rss_mb": 311.1,
   "compile_s": null
  },
  "safec_heap": {
   "release": 0.2291,
   "rss_mb": 200.4,
   "compile_s": 0.048
  }
 },
 "simd": {
  "safec": {
   "scalar": 0.0799,
   "vec": 0.0747
  },
  "c": {
   "scalar": 0.0802,
   "vec": 0.077
  },
  "cpp": {
   "scalar": 0.0806,
   "vec": 0.0739
  },
  "rust": {
   "scalar": 0.0804,
   "vec": 0.0767
  },
  "zig": {
   "scalar": 0.0786,
   "vec": 0.072
  },
  "go": {
   "scalar": 0.0853
  },
  "python": {
   "scalar": 3.1168,
   "vec": 0.1857
  }
 },
 "collections": {
  "safec": {
   "vec push": 2.307,
   "vec sort": 43.455,
   "map insert": 37.884,
   "map get": 8.358,
   "list push+pop": 10.071,
   "queue enq+deq": 3.025,
   "bst insert": 175.608,
   "bst get": 173.676
  },
  "c": {
   "vec push": 1.51,
   "vec sort": 83.416,
   "map insert": 34.343,
   "map get": 7.385,
   "list push+pop": 22.245,
   "queue enq+deq": 5.535,
   "bst insert": 179.936,
   "bst get": 172.654
  }
 },
 "memalloc": {
  "std_ns_per_pair": 3.57,
  "malloc_ns_per_pair": 6.38,
  "std_ms": 87.15,
  "malloc_ms": 96.22
 },
 "ml": {
  "small/safec_cpu": {
   ":train_ms": 16.808,
   ":last_loss": 100.333267,
   ":inference_ms_per_1000": 36.027
  },
  "small/safec_blas": {
   ":train_ms": 9.569,
   ":last_loss": 100.333267,
   ":inference_ms_per_1000": 24.272
  },
  "big/safec_cpu": {
   ":train_ms": 48.774,
   ":last_loss": 95.426926,
   ":inference_ms_per_200": 66.967
  },
  "big/safec_blas": {
   ":train_ms": 93.35,
   ":last_loss": 95.424927,
   ":inference_ms_per_200": 123.11
  },
  "small/safec_cpu_aocl": {
   ":train_ms": 14.852,
   ":last_loss": 100.333267,
   ":inference_ms_per_1000": 26.27
  },
  "big/safec_cpu_aocl": {
   ":train_ms": 36.518,
   ":last_loss": 95.427536,
   ":inference_ms_per_200": 43.645
  },
  "small/safec_cpu_mkl": {
   ":train_ms": 18.088,
   ":last_loss": 100.333267,
   ":inference_ms_per_1000": 31.384
  },
  "big/safec_cpu_mkl": {
   ":train_ms": 45.403,
   ":last_loss": 95.428177,
   ":inference_ms_per_200": 52.58
  },
  "rocm/safec": {
   "small 128-256-64 plain": {
    "train_ms": 6.901,
    "last_loss": 100.333275,
    "throughput_samples_per_sec": 921796.53,
    "inference_ms_per_1000": 33.496,
    "checksum": -98.959446
   },
   "big 512-1024-256 plain": {
    "train_ms": 7.767,
    "last_loss": 95.426941,
    "throughput_samples_per_sec": 824003.22,
    "inference_ms_per_200": 10.886,
    "checksum": 0.730239
   },
   "small 128-256-64 graph": {
    "train_ms": 6.718,
    "last_loss": 100.33329,
    "throughput_samples_per_sec": 938043.25,
    "inference_ms_per_1000": 36.911,
    "checksum": -98.959446
   },
   "big 512-1024-256 graph": {
    "train_ms": 7.576,
    "last_loss": 95.426956,
    "throughput_samples_per_sec": 843527.83,
    "inference_ms_per_200": 11.509,
    "checksum": 0.730239
   },
   "small 128-256-64 fp16": {
    "train_ms": 5.534,
    "last_loss": 100.336121,
    "throughput_samples_per_sec": 1057825.53,
    "inference_ms_per_1000": 33.833,
    "checksum": -98.937988
   },
   "big 512-1024-256 fp16": {
    "train_ms": 5.929,
    "last_loss": 93.010933,
    "throughput_samples_per_sec": 1063475.37,
    "inference_ms_per_200": 9.259,
    "checksum": 0.795746
   },
   "small 128-256-64 bf16": {
    "train_ms": 5.783,
    "last_loss": 99.90625,
    "throughput_samples_per_sec": 1061862.98,
    "inference_ms_per_1000": 37.279,
    "checksum": -90.332031
   },
   "big 512-1024-256 bf16": {
    "train_ms": 6.093,
    "last_loss": 63.433678,
    "throughput_samples_per_sec": 1050406.37,
    "inference_ms_per_200": 9.865,
    "checksum": -2.246094
   }
  },
  "small/pytorch_cpu": {
   "cpu:train_ms": 14.038,
   ":last_loss": 100.018272,
   "cpu:inference_ms_per_1000": 21.88
  },
  "small/pytorch_cuda": {
   "cuda:train_ms": 19.632,
   ":last_loss": 103.071129,
   "cuda:inference_ms_per_1000": 36.219
  },
  "small/pytorch_cuda_fp16": {
   "cuda:train_ms": 19.13,
   ":last_loss": 103.0625,
   "cuda:inference_ms_per_1000": 33.716
  },
  "small/pytorch_cuda_bf16": {
   "cuda:train_ms": 20.411,
   ":last_loss": 103.0,
   "cuda:inference_ms_per_1000": 47.064
  },
  "big/pytorch_cpu": {
   "cpu:train_ms": 38.353,
   ":last_loss": 106.601181,
   "cpu:inference_ms_per_200": 50.832
  },
  "big/pytorch_cuda": {
   "cuda:train_ms": 10.105,
   ":last_loss": 1868.155273,
   "cuda:inference_ms_per_200": 13.348
  },
  "big/pytorch_cuda_fp16": {
   "cuda:train_ms": 9.902,
   ":last_loss": 1942.0,
   "cuda:inference_ms_per_200": 17.864
  },
  "big/pytorch_cuda_bf16": {
   "cuda:train_ms": 11.797,
   ":last_loss": 107.0,
   "cuda:inference_ms_per_200": 15.882
  },
  "small/tensorflow_cpu": {
   "cpu:train_ms": 79.527,
   ":last_loss": 106.182999,
   "cpu:inference_ms_per_1000": 203.889
  },
  "big/tensorflow_cpu": {
   "cpu:train_ms": 105.759,
   ":last_loss": 316.611603,
   "cpu:inference_ms_per_200": 123.564
  }
 },
 "web": {
  "safec": {
   "rps": 25302.88,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 4.9
  },
  "c": {
   "rps": 24620.48,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 3.1
  },
  "cpp": {
   "rps": 24844.97,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 5.1
  },
  "zig": {
   "rps": 25030.1,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 7.1
  },
  "go": {
   "rps": 23198.76,
   "p50_ms": 2,
   "p99_ms": 3,
   "failed": 0,
   "rss_mb": 16.1
  },
  "rust": {
   "rps": 25074.88,
   "p50_ms": 2,
   "p99_ms": 2,
   "failed": 0,
   "rss_mb": 4.3
  },
  "python": {
   "rps": 5095.51,
   "p50_ms": 10,
   "p99_ms": 14,
   "failed": 0,
   "rss_mb": 46.0
  }
 },
 "conv": {
  "safec_cpu": {
   "conv64 b=1 cpu": {
    "ms_per_image": 0.412,
    "checksum": 528.229
   },
   "unet128 b=1 cpu": {
    "ms_per_image": 1.684,
    "checksum": 994785.1153
   }
  },
  "safec_cpu_aocl": {
   "conv64 b=1 cpu": {
    "ms_per_image": 0.359,
    "checksum": 528.229
   },
   "unet128 b=1 cpu": {
    "ms_per_image": 1.352,
    "checksum": 994785.1012
   }
  },
  "safec_cpu_mkl": {
   "conv64 b=1 cpu": {
    "ms_per_image": 0.399,
    "checksum": 528.229
   },
   "unet128 b=1 cpu": {
    "ms_per_image": 1.691,
    "checksum": 994785.0986
   }
  },
  "safec_rocm": {
   "conv64 b=1 fp32": {
    "ms_per_batch": 0.027,
    "ms_per_image": 0.02697
   },
   "conv64 b=32 fp32": {
    "ms_per_batch": 0.3423,
    "ms_per_image": 0.0107
   },
   "unet128 b=1 fp32": {
    "ms_per_batch": 0.2217,
    "ms_per_image": 0.22172
   },
   "unet128 b=1 graph fp32": {
    "ms_per_batch": 0.2163,
    "ms_per_image": 0.2163
   },
   "unet128 b=16 fp32": {
    "ms_per_batch": 1.2892,
    "ms_per_image": 0.08058
   }
  },
  "pytorch_cpu": {
   "conv64 b=1 cpu": {
    "ms_per_batch": 0.3265,
    "ms_per_image": 0.32652
   },
   "unet128 b=1 cpu": {
    "ms_per_batch": 2.6755,
    "ms_per_image": 2.67553
   }
  },
  "pytorch_cuda": {
   "conv64 b=1 fp32": {
    "ms_per_batch": 0.0628,
    "ms_per_image": 0.06284
   },
   "conv64 b=32 fp32": {
    "ms_per_batch": 0.3463,
    "ms_per_image": 0.01082
   },
   "unet128 b=1 fp32": {
    "ms_per_batch": 0.3144,
    "ms_per_image": 0.31445
   },
   "unet128 b=16 fp32": {
    "ms_per_batch": 1.2588,
    "ms_per_image": 0.07867
   },
   "conv64 b=1 tf32": {
    "ms_per_batch": 0.0628,
    "ms_per_image": 0.06277
   },
   "conv64 b=32 tf32": {
    "ms_per_batch": 0.3478,
    "ms_per_image": 0.01087
   },
   "unet128 b=1 tf32": {
    "ms_per_batch": 0.3135,
    "ms_per_image": 0.31346
   },
   "unet128 b=16 tf32": {
    "ms_per_batch": 1.2598,
    "ms_per_image": 0.07874
   },
   "conv64 b=1 fp32 heuristic": {
    "ms_per_batch": 0.0628,
    "ms_per_image": 0.06277
   },
   "conv64 b=32 fp32 heuristic": {
    "ms_per_batch": 0.3483,
    "ms_per_image": 0.01088
   },
   "unet128 b=1 fp32 heuristic": {
    "ms_per_batch": 0.3143,
    "ms_per_image": 0.31426
   },
   "unet128 b=16 fp32 heuristic": {
    "ms_per_batch": 1.2605,
    "ms_per_image": 0.07878
   }
  }
 },
 "fft": {
  "safec_cpu": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 0.888
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 0.691
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 1.255
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 0.861
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 0.688
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 0.56
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 0.775
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 0.558
   }
  },
  "safec_cpu_aocl": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 0.916
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 0.691
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 1.221
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 0.852
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 0.662
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 0.546
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 0.776
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 0.559
   }
  },
  "safec_cpu_mkl": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 0.891
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 1.001
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 1.888
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 1.049
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 0.693
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 0.543
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 0.964
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 0.457
   }
  },
  "safec_rocm": {
   "fft1d n=1048576 c128 rocm": {
    "ms_per_call": 1.4607
   },
   "fft2d 1024x1024 c128 rocm": {
    "ms_per_call": 1.428
   },
   "batch 1024x4096 c2c rocm": {
    "ms_per_call": 0.1002
   },
   "rfft 1024x4096 r2c rocm": {
    "ms_per_call": 0.0949
   }
  },
  "pytorch_cpu": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 8.3101
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 0.7101
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 4.412
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 0.4428
   }
  },
  "pytorch_cuda": {
   "fft1d n=1048576 c128 cuda": {
    "ms_per_call": 1.5563
   },
   "fft2d 1024x1024 c128 cuda": {
    "ms_per_call": 1.5282
   },
   "batch 1024x4096 c2c cuda": {
    "ms_per_call": 0.1709
   },
   "rfft 1024x4096 r2c cuda": {
    "ms_per_call": 0.1264
   }
  },
  "fftw_1t": {
   "fft1d n=1048576 c128 st": {
    "ms_per_call": 6.505
   },
   "fft2d 1024x1024 c128 st": {
    "ms_per_call": 3.488
   },
   "batch 1024x4096 c2c st": {
    "ms_per_call": 7.839
   },
   "rfft 1024x4096 r2c st": {
    "ms_per_call": 4.275
   },
   "fft1d n=1048576 c64 st": {
    "ms_per_call": 5.529
   },
   "fft2d 1024x1024 c64 st": {
    "ms_per_call": 2.541
   },
   "batch 1024x4096 c2c f32 st": {
    "ms_per_call": 3.311
   },
   "rfft 1024x4096 r2c f32 st": {
    "ms_per_call": 2.978
   }
  },
  "fftw_16t": {
   "fft1d n=1048576 c128 mt": {
    "ms_per_call": 1.629
   },
   "fft2d 1024x1024 c128 mt": {
    "ms_per_call": 0.618
   },
   "batch 1024x4096 c2c mt": {
    "ms_per_call": 1.053
   },
   "rfft 1024x4096 r2c mt": {
    "ms_per_call": 0.553
   },
   "fft1d n=1048576 c64 mt": {
    "ms_per_call": 1.673
   },
   "fft2d 1024x1024 c64 mt": {
    "ms_per_call": 0.446
   },
   "batch 1024x4096 c2c f32 mt": {
    "ms_per_call": 0.504
   },
   "rfft 1024x4096 r2c f32 mt": {
    "ms_per_call": 0.409
   }
  },
  "kfr": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 3.359
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 4.217
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 11.81
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 4.342
   },
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 31.907
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 4.863
   },
   "biquad n=1048576": {
    "ms_per_call": 2.342
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 62.458
   },
   "resample draft n=1048576 44100->48000": {
    "ms_per_call": 26.281
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 1.766
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 2.63
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 4.226
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 2.13
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 27.754
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 2.75
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 2.148
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 34.204
   },
   "resample draft n=1048576 44100->48000 f32": {
    "ms_per_call": 27.271
   }
  },
  "rust": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 4.751
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 8.251
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 2.102
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 4.105
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 525.809
   }
  }
 },
 "dsp": {
  "safec_cpu": {
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 5.868,
    "checksum": 965.503541
   },
   "fir block n=1048576 taps=127": {
    "ms_per_call": 0.841,
    "checksum": 965.503541
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 2.981,
    "checksum": 11438.877046
   },
   "biquad n=1048576": {
    "ms_per_call": 0.484,
    "checksum": -646.722217
   },
   "stft n=1048576 frame=1024 hop=256": {
    "ms_per_call": 1.824,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 3.711,
    "checksum": -704.289099
   },
   "conv2d 1024x1024 k=5x5": {
    "ms_per_call": 0.366,
    "checksum": 205.652439
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 3.035,
    "checksum": 965.503677
   },
   "fir block n=1048576 taps=127 f32": {
    "ms_per_call": 0.475,
    "checksum": 965.503685
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 1.286,
    "checksum": 11438.881763
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 0.482,
    "checksum": -646.722416
   },
   "stft n=1048576 frame=1024 hop=256 f32": {
    "ms_per_call": 0.986,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 3.957,
    "checksum": -704.289077
   }
  },
  "safec_cpu_aocl": {
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 5.974,
    "checksum": 965.503541
   },
   "fir block n=1048576 taps=127": {
    "ms_per_call": 0.824,
    "checksum": 965.503541
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 3.021,
    "checksum": 11438.877046
   },
   "biquad n=1048576": {
    "ms_per_call": 0.484,
    "checksum": -646.722217
   },
   "stft n=1048576 frame=1024 hop=256": {
    "ms_per_call": 1.85,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 3.838,
    "checksum": -704.289099
   },
   "conv2d 1024x1024 k=5x5": {
    "ms_per_call": 0.369,
    "checksum": 205.652439
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 3.026,
    "checksum": 965.503677
   },
   "fir block n=1048576 taps=127 f32": {
    "ms_per_call": 0.45,
    "checksum": 965.503685
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 1.307,
    "checksum": 11438.881763
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 0.483,
    "checksum": -646.722416
   },
   "stft n=1048576 frame=1024 hop=256 f32": {
    "ms_per_call": 0.968,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 3.819,
    "checksum": -704.289077
   }
  },
  "safec_cpu_mkl": {
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 5.858,
    "checksum": 965.503541
   },
   "fir block n=1048576 taps=127": {
    "ms_per_call": 0.796,
    "checksum": 965.503541
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 3.191,
    "checksum": 11438.877046
   },
   "biquad n=1048576": {
    "ms_per_call": 0.493,
    "checksum": -646.722217
   },
   "stft n=1048576 frame=1024 hop=256": {
    "ms_per_call": 1.815,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 3.878,
    "checksum": -704.289099
   },
   "conv2d 1024x1024 k=5x5": {
    "ms_per_call": 0.39,
    "checksum": 205.652439
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 3.069,
    "checksum": 965.503677
   },
   "fir block n=1048576 taps=127 f32": {
    "ms_per_call": 0.471,
    "checksum": 965.503685
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 1.162,
    "checksum": 11438.875078
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 0.525,
    "checksum": -646.722416
   },
   "stft n=1048576 frame=1024 hop=256 f32": {
    "ms_per_call": 0.983,
    "frames": 4093.0
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 3.98,
    "checksum": -704.289077
   }
  },
  "scipy": {
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 17.416
   },
   "fir block n=1048576 taps=127": {
    "ms_per_call": 15.441
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 32.303
   },
   "biquad n=1048576": {
    "ms_per_call": 4.708
   },
   "stft n=1048576 frame=1024 hop=256": {
    "ms_per_call": 53.946
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 9.07
   },
   "conv2d 1024x1024 k=5x5": {
    "ms_per_call": 35.322
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 19.529
   },
   "fir block n=1048576 taps=127 f32": {
    "ms_per_call": 18.886
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 14.783
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 4.595
   },
   "stft n=1048576 frame=1024 hop=256 f32": {
    "ms_per_call": 55.759
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 8.684
   }
  },
  "kfr": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 3.386
   },
   "fft2d 1024x1024 c128 cpu": {
    "ms_per_call": 4.28
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 11.888
   },
   "rfft 1024x4096 r2c cpu": {
    "ms_per_call": 4.274
   },
   "fir stream n=1048576 taps=127": {
    "ms_per_call": 32.051
   },
   "fftconv n=1048576 taps=4096": {
    "ms_per_call": 4.89
   },
   "biquad n=1048576": {
    "ms_per_call": 2.327
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 61.721
   },
   "resample draft n=1048576 44100->48000": {
    "ms_per_call": 26.277
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 1.693
   },
   "fft2d 1024x1024 c64 cpu": {
    "ms_per_call": 2.679
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 4.358
   },
   "rfft 1024x4096 r2c f32 cpu": {
    "ms_per_call": 2.127
   },
   "fir stream n=1048576 taps=127 f32": {
    "ms_per_call": 27.817
   },
   "fftconv n=1048576 taps=4096 f32": {
    "ms_per_call": 2.759
   },
   "biquad n=1048576 f32": {
    "ms_per_call": 2.152
   },
   "resample n=1048576 44100->48000 f32": {
    "ms_per_call": 34.121
   },
   "resample draft n=1048576 44100->48000 f32": {
    "ms_per_call": 27.254
   }
  },
  "q_madronalib": {
   "biquad q n=1048576 f32": {
    "ms_per_call": 1.69,
    "checksum": 996.364585
   },
   "biquad madronalib n=1048576 f32": {
    "ms_per_call": 2.874,
    "checksum": 996.365938
   }
  },
  "rust": {
   "fft1d n=1048576 c128 cpu": {
    "ms_per_call": 4.751
   },
   "batch 1024x4096 c2c cpu": {
    "ms_per_call": 8.29
   },
   "fft1d n=1048576 c64 cpu": {
    "ms_per_call": 2.109
   },
   "batch 1024x4096 c2c f32 cpu": {
    "ms_per_call": 4.074
   },
   "resample n=1048576 44100->48000": {
    "ms_per_call": 519.222
   }
  }
 },
 "attention": {
  "safec_rocm": {
   "flash fp32 d64 s1024 c0": {
    "ms_per_call": 0.8574,
    "tflops": 19.99
   },
   "flash fp32 d64 s1024 c1": {
    "ms_per_call": 0.4567,
    "tflops": 18.74
   },
   "flash fp32 d64 s4096 c0": {
    "ms_per_call": 10.6679,
    "tflops": 25.77
   },
   "flash fp32 d64 s4096 c1": {
    "ms_per_call": 5.4755,
    "tflops": 25.06
   },
   "flash fp32 d128 s1024 c0": {
    "ms_per_call": 1.8069,
    "tflops": 18.83
   },
   "flash fp32 d128 s1024 c1": {
    "ms_per_call": 0.9513,
    "tflops": 18.06
   },
   "flash fp32 d128 s4096 c0": {
    "ms_per_call": 24.6872,
    "tflops": 22.22
   },
   "flash fp32 d128 s4096 c1": {
    "ms_per_call": 12.5527,
    "tflops": 21.88
   },
   "linear fp32 d64 s4096 c0": {
    "ms_per_call": 0.4638
   },
   "linear fp32 d64 s4096 c1": {
    "ms_per_call": 0.6729
   },
   "linear fp32 d64 s16384 c0": {
    "ms_per_call": 1.6129
   },
   "linear fp32 d64 s16384 c1": {
    "ms_per_call": 2.7556
   },
   "linear fp32 d128 s4096 c0": {
    "ms_per_call": 1.5977
   },
   "linear fp32 d128 s4096 c1": {
    "ms_per_call": 2.8428
   }
  },
  "pytorch_rocm": {
   "flash fp32 d64 s1024 c0": {
    "ms_per_call": 2.5167,
    "tflops": 6.82
   },
   "flash fp32 d64 s1024 c1": {
    "ms_per_call": 3.2778,
    "tflops": 2.62
   },
   "flash fp32 d64 s4096 c0": {
    "ms_per_call": 42.8119,
    "tflops": 6.42
   },
   "flash fp32 d64 s4096 c1": {
    "ms_per_call": 56.1445,
    "tflops": 2.44
   },
   "flash fp32 d128 s1024 c0": {
    "ms_per_call": 3.3144,
    "tflops": 10.34
   },
   "flash fp32 d128 s1024 c1": {
    "ms_per_call": 4.0469,
    "tflops": 4.24
   },
   "flash fp32 d128 s4096 c0": {
    "ms_per_call": 53.4555,
    "tflops": 10.27
   },
   "flash fp32 d128 s4096 c1": {
    "ms_per_call": 66.3687,
    "tflops": 4.13
   },
   "linear fp32 d64 s4096 c0": {
    "ms_per_call": 1.3999
   },
   "linear fp32 d64 s4096 c1": {
    "ms_per_call": 3.6416
   },
   "linear fp32 d64 s16384 c0": {
    "ms_per_call": 5.6367
   },
   "linear fp32 d64 s16384 c1": {
    "ms_per_call": 14.5582
   },
   "linear fp32 d128 s4096 c0": {
    "ms_per_call": 3.2565
   },
   "linear fp32 d128 s4096 c1": {
    "ms_per_call": 8.7007
   }
  }
 },
 "ml_nn": {
  "safec_cpu": {
   "mlp_small": {
    "train_ms": 3.85062,
    "last_loss": 0.0297225
   },
   "mlp_big": {
    "train_ms": 21.1123,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 15.5828,
    "last_loss": 31.5691
   },
   "gan": {
    "train_ms": 9.03194,
    "last_loss": 1.40552
   },
   "dqn": {
    "train_ms": 52.3699,
    "last_loss": 0.138698
   },
   "ppo": {
    "train_ms": 27.0087,
    "last_loss": 48.7648
   },
   "lora": {
    "train_ms": 11.7603,
    "last_loss": 0.0219908
   },
   "conv_ae": {
    "train_ms": 30.5869,
    "last_loss": 0.255294
   }
  },
  "safec_cpu_aocl": {
   "mlp_small": {
    "train_ms": 3.64135,
    "last_loss": 0.0297226
   },
   "mlp_big": {
    "train_ms": 19.9322,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 14.7443,
    "last_loss": 31.5691
   },
   "gan": {
    "train_ms": 8.6152,
    "last_loss": 1.40552
   },
   "dqn": {
    "train_ms": 51.2142,
    "last_loss": 0.138698
   },
   "ppo": {
    "train_ms": 25.5483,
    "last_loss": 48.7648
   },
   "lora": {
    "train_ms": 11.2142,
    "last_loss": 0.0219908
   },
   "conv_ae": {
    "train_ms": 30.6631,
    "last_loss": 0.255294
   }
  },
  "safec_cpu_mkl": {
   "mlp_small": {
    "train_ms": 3.80954,
    "last_loss": 0.0297225
   },
   "mlp_big": {
    "train_ms": 23.5696,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 15.1532,
    "last_loss": 31.5691
   },
   "gan": {
    "train_ms": 9.11665,
    "last_loss": 1.40552
   },
   "dqn": {
    "train_ms": 51.7616,
    "last_loss": 0.138698
   },
   "ppo": {
    "train_ms": 25.7478,
    "last_loss": 48.7648
   },
   "lora": {
    "train_ms": 11.8873,
    "last_loss": 0.0219908
   },
   "conv_ae": {
    "train_ms": 30.4276,
    "last_loss": 0.255294
   }
  },
  "safec_rocm": {
   "mlp_small": {
    "train_ms": 20.0526,
    "last_loss": 0.0297226
   },
   "mlp_big": {
    "train_ms": 6.12657,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 12.6438,
    "last_loss": 31.5345
   },
   "gan": {
    "train_ms": 26.8618,
    "last_loss": 1.40854
   },
   "dqn": {
    "train_ms": 155.567,
    "last_loss": 0.138698
   },
   "ppo": {
    "train_ms": 37.3321,
    "last_loss": 48.7648
   },
   "conv_ae": {
    "train_ms": 12.0606,
    "last_loss": 0.255294
   }
  },
  "pytorch_cpu": {
   "mlp_small": {
    "train_ms": 158.895,
    "last_loss": 0.0331768
   },
   "mlp_big": {
    "train_ms": 35.929,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 97.02,
    "last_loss": 50.6271
   },
   "gan": {
    "train_ms": 144.424,
    "last_loss": 1.42086
   },
   "dqn": {
    "train_ms": 624.455,
    "last_loss": 0.126866
   },
   "ppo": {
    "train_ms": 267.649,
    "last_loss": 41.1927
   },
   "lora": {
    "train_ms": 209.768,
    "last_loss": 0.0405097
   },
   "conv_ae": {
    "train_ms": 88.945,
    "last_loss": 0.484414
   }
  },
  "pytorch_cuda": {
   "mlp_small": {
    "train_ms": 245.513,
    "last_loss": 0.0559914
   },
   "mlp_big": {
    "train_ms": 31.808,
    "last_loss": 0.0
   },
   "vae": {
    "train_ms": 138.481,
    "last_loss": 50.4153
   },
   "gan": {
    "train_ms": 248.9,
    "last_loss": 1.47774
   },
   "dqn": {
    "train_ms": 1167.515,
    "last_loss": 0.223382
   },
   "ppo": {
    "train_ms": 382.785,
    "last_loss": 36.4086
   },
   "lora": {
    "train_ms": 355.591,
    "last_loss": 0.0436069
   },
   "conv_ae": {
    "train_ms": 80.324,
    "last_loss": 0.50067
   }
  }
 },
 "ml_nn_gpu": {
  "safec_cpu": {
   "mlp": {
    "train_ms": 205.205,
    "last_loss": 0.0
   },
   "cnn": {
    "train_ms": 863.693,
    "last_loss": 0.080832
   },
   "resnet": {
    "train_ms": 2850.26,
    "last_loss": 0.13114
   },
   "unet": {
    "train_ms": 2343.6,
    "last_loss": 0.137156
   },
   "dcgan": {
    "train_ms": 763.91,
    "last_loss": 0.443979
   },
   "convvae": {
    "train_ms": 148.802,
    "last_loss": 393.525
   }
  },
  "safec_rocm": {
   "mlp_fp32": {
    "train_ms": 27.625,
    "last_loss": 0.0
   },
   "cnn_fp32": {
    "train_ms": 31.2095,
    "last_loss": 0.080832
   },
   "resnet_fp32": {
    "train_ms": 75.3022,
    "last_loss": 0.131137
   },
   "unet_fp32": {
    "train_ms": 50.547,
    "last_loss": 0.137156
   },
   "dcgan_fp32": {
    "train_ms": 56.0834,
    "last_loss": 0.454596
   },
   "convvae_fp32": {
    "train_ms": 26.769,
    "last_loss": 391.253
   },
   "mlp_fp16": {
    "train_ms": 26.6415,
    "last_loss": 0.0
   },
   "cnn_fp16": {
    "train_ms": 29.6268,
    "last_loss": 0.0807998
   },
   "resnet_fp16": {
    "train_ms": 57.9073,
    "last_loss": 0.131178
   },
   "unet_fp16": {
    "train_ms": 44.6985,
    "last_loss": 0.137455
   },
   "mlp_bf16": {
    "train_ms": 25.6857,
    "last_loss": 0.0
   },
   "cnn_bf16": {
    "train_ms": 32.4877,
    "last_loss": 0.0812737
   },
   "resnet_bf16": {
    "train_ms": 79.5407,
    "last_loss": 0.131495
   },
   "unet_bf16": {
    "train_ms": 69.8064,
    "last_loss": 0.137935
   }
  },
  "pytorch_cuda": {
   "mlp_fp32": {
    "train_ms": 56.586,
    "last_loss": 0.0
   },
   "cnn_fp32": {
    "train_ms": 60.683,
    "last_loss": 1.38697
   },
   "resnet_fp32": {
    "train_ms": 98.892,
    "last_loss": 0.137842
   },
   "unet_fp32": {
    "train_ms": 146.361,
    "last_loss": 0.396237
   },
   "dcgan_fp32": {
    "train_ms": 106.197,
    "last_loss": 0.771293
   },
   "convvae_fp32": {
    "train_ms": 53.261,
    "last_loss": 444.263
   },
   "mlp_fp16": {
    "train_ms": 70.172,
    "last_loss": 0.0
   },
   "cnn_fp16": {
    "train_ms": 58.312,
    "last_loss": 1.40338
   },
   "resnet_fp16": {
    "train_ms": 74.343,
    "last_loss": 2.31742
   },
   "unet_fp16": {
    "train_ms": 153.543,
    "last_loss": 0.416703
   },
   "mlp_bf16": {
    "train_ms": 63.301,
    "last_loss": 0.0
   },
   "cnn_bf16": {
    "train_ms": 94.001,
    "last_loss": 1.39249
   },
   "resnet_bf16": {
    "train_ms": 271.435,
    "last_loss": 0.1593
   },
   "unet_bf16": {
    "train_ms": 276.962,
    "last_loss": 0.410217
   }
  }
 }
}