Same-code process calibration, 2026-10-18

The fixed same-code calibration completed all seven configurations and retained 112 fresh-worker artifacts, 1,408 eligible candidate/case outcomes and 88 primary comparison rows. Of those rows, 33 supported no material regression within the original 5% budget, 55 were inconclusive and none had a lower confidence bound above 1.05. All four campaign-level verdicts remained inconclusive.

This measures the procedure’s ability to distinguish two aliases of the same build on this host. It does not measure an optimization, establish a speedup, or establish that the host will remain stable. In particular, zero strict regressions does not turn the 55 inconclusive rows into passes. The A/B labels below are baseline/current aliases of identical payloads; a NumPy row compares NumPy with itself, and a PyTorch row compares PyTorch with itself.

The earlier Complete current-main CPU and CUDA baseline, 2026-10-17 framework baseline and Current-main fixed A/B optimization evidence — 2026-10-17 rejected trials retain their original sources, protocols, outcomes and URLs. These observations are a separate calibration; no earlier trial was replaced, combined with this one or corrected using a framework control.

Protocol and sources

The solver packages were compiled from commit 84c88c98c895103395e1b6d6e98f7618848a14d3, tree 57a755d537e516b91c70c1b555f2803753389862. Within each preset, baseline and current refer to the exact same frozen package root and file manifest, including the native extension. Presets have separate builds; they are not aliases of each other.

The measured orchestration source is commit a4655b94c0893e6b5f68a102ea4730292baab074, tree e8f521211627ea1b2798ceef50cc48ff25b3bac1. Its frozen runner SHA256 is ca9844ab88a5fe99c8a05cc1403786e2492ee746bbf43aa78d0c47831ae7c654. A later failure-cleanup repair has separate source and validation; it does not change the successful launch commands used here. These measurements must not be relabeled as measurements of that later runner revision.

Each of two time windows runs ABBA followed by BAAB for every configuration. Window 2 reverses configuration order. Every slot launches the public python -m cgbench run interface with a fresh timing worker. This gives eight artifact blocks per side, each with three warmups, 36 complete Williams rounds and one invocation per round: 288 normalized timing rounds per side and row. The windows have a declared 60-second pause; that pause does not establish thermal equilibrium or independence of the physical host.

Fixture seed is 20261017. Every configuration uses float64, rtol=1e-8, atol=1e-12, maxiter=2000 and residual_refresh=50. Ordered cases are dense-64, dense-256, dense-1024 and dense-2048 from the shared cpu-dense-f64-v1 / cuda-dense-f64-v1 fixtures. Input hashes, actual iterations, matvecs and known restarts match across all same-candidate A/B blocks. Every retained outcome passed the independent true-residual gate. No work normalization, row exclusions, selective retries or control-drift subtraction/division was applied.

The original, unchanged comparator has SHA256 b057d071ebe47e8356a38cc2c43624ea15ff13a34abbcb02b249c83c52f23063. It uses 10,000 artifact-then-round hierarchical bootstrap samples, seed 20261018, inclusive percentile 95% intervals and the fixed 5% budget. An upper bound at most 1.05 supports no material regression; a lower bound above 1.05 indicates regression; overlap is inconclusive. Every native and unchanged framework control participates in the same primary decision. A supported row is not evidence of a statistically established speedup.

ID

Preset

Scope

Algorithm

M

Fusion

Native mode

Threads / logical CPUs

C1

cpu-release

prepared

standard

none

on

host

1 / 0

C2

cpu-blas

prepared

explicit

none

off

host

1 / 0

C3

cpu-openmp

prepared

standard

jacobi

on

host

2 / 0,2

C4

cuda-release

api

standard

jacobi

on

host

1 / 0

C5

cuda-release

api

standard

jacobi

on

device

1 / 0

C6

cuda-release

session

standard

jacobi

on

host

1 / 0

C7

cuda-release

prepared

standard

jacobi

on

device

1 / 0

C1 and C3 retain native,numpy-out,torch-eager-out. C2 retains native,native-blas,numpy-out,torch-eager-out. C4–C7 retain cuda-native,cuda-cublas,torch-cuda-eager. The native CUDA mode selection is recorded as inapplicable to the eager PyTorch reference; host/device do not select a different PyTorch engine. There is no CPU Session measurement.

Fresh CUDA API includes output, workspace and per-call Jacobi setup; resident input uploads are outside the timer. Session reuses the iterative workspace while charging fresh output and per-call Jacobi. Prepared reuses output, workspace and preconditioner; each measured invocation still resets x0. Completed-wall fences remain in the declared timing protocol, while correctness downloads remain outside the measured intervals. Framework allocator/handle caches keep their normal behavior; fresh API does not mean a cold framework runtime on every solve. No hidden native reuse or AUTO selection is introduced.

Observed environment and limitations

The host reports an Intel Core i7-13700K, 24 logical CPUs and Linux 6.6.87.2 under WSL2. Python is 3.12.3, NumPy 2.2.6, SciPy 1.15.3 and threadpoolctl 3.6.0. CPU campaigns use PyTorch 2.8.0+cpu; the CUDA campaign uses PyTorch 2.8.0+cu128. CPU native builds report GNU 16.0.1; the CUDA host build reports GNU 13.3.0 and NVIDIA CUDA compiler 12.8.93, architecture 75. Native Release flags are -O3 -DNDEBUG without fast math or host ISA tuning. The explicit BLAS build uses the private, sequential LP64 OpenBLAS 0.3.30 provider with cgpy_ symbols; NumPy reports its separate OpenBLAS 0.3.29 provider. Provider/build inventories remain in the raw artifacts.

The GPU is an NVIDIA GeForce RTX 2060, compute capability 7.5, UUID GPU-725be96d-2b67-c8ec-fa10-5d9b23a2c302. The driver reports 610.88; CUDA runtime is 12080 and driver API version 13030. CUDA controls are explicitly CUBLAS_WORKSPACE_CONFIG=:4096:8 and no CUDA_LAUNCH_BLOCKING. All six thread-pool environment variables, observed pool counts, Torch intraop/interop counts and logical CPU affinity were checked. OMP_DYNAMIC and MKL_DYNAMIC are false. Logical CPU affinity does not identify isolated physical cores or exclude Windows activity.

Boundary telemetry records UTC and monotonic timestamps, Linux load/stat/CPU pressure, available sensors and read-only nvidia-smi output before and after each child, outside solve timers. CPU sysfs frequency and temperature sensors are unavailable in these captures. /proc/cpuinfo MHz strings do not establish the physical CPU clock during a solve. GPU query values are boundary observations, not a continuous account of clocks, utilization, thermal state or another application’s activity. Telemetry is diagnostic; its own driver queries can perturb the system.

One additional manual read-only GPU utilization/temperature query occurred during the CUDA calibration. It reported 23, 50. The last completed slot observed before that query was 43; the latest slot recorded at disclosure was 60. Exact query start/end timestamps were not instrumented. The event was recorded at 2026-10-17T16:48:14.175741+00:00 in additional-observer-query.json; that later recording time is not its execution timestamp. Every row and block remains present, without exclusions, correction, rerun or pooling. This extra query is not evidence of an isolated or continuously monitored host.

Eight fresh artifact blocks per side provide more process observations than the earlier two-block trials, but only two declared time windows were observed. Fresh processes do not prove independent environmental states. The unchanged primary comparator resamples rounds independently within each artifact; serial dependence can make intervals too narrow. Intervals are per row, without simultaneous-coverage adjustment. WSL does not isolate Windows host load, display activity, power management or thermal behavior. These limits apply even where the 5% gate is supported. This calibration does not quantify an improvement in false-regression rate against earlier trials from another time window, and cannot attribute a timing shift to a specific disturbance.

Before calibration, an initial CPU setup test failed because the inherited CG_REQUIRE_CUDA=1 requirement was applied to a CPU-only build. Its original log and summary remain unchanged. The corrected setup cleared that requirement for CPU presets and enabled it only for CUDA. All four corrected build/test commands succeeded before timing began. This was a setup correction, not a retrial or exclusion of a started measurement campaign.

Complete primary outcomes

Preset

Configs

Artifacts

Eligible outcomes

Rows

Within budget

Inconclusive

Regression

cpu-release

1

16

192

12

0

12

0

cpu-blas

1

16

256

16

5

11

0

cpu-openmp

1

16

192

12

3

9

0

cuda-release

4

64

768

48

25

23

0

Total

7

112

1408

88

33

55

0

All four campaign manifests have status=complete and verdict=inconclusive. Their command exit status is 1 because inconclusive is not a supported pass; it is not a worker crash or numerical failure. The observations were collected between 2026-10-18 00:28:35 and 00:48:50 in Asia/Taipei (2026-10-17 16:28:35–16:48:50 UTC).

In the following tables, A/B milliseconds are pooled medians over all 288 normalized rounds on that side, not medians of the eight process medians. W means supported_no_material_regression and I means inconclusive. I/M/R gives the verified iteration/matvec/restart counts for that same candidate on both aliases. B/A is the comparator’s original ratio of pooled medians. Displayed decimal values are rounded; the reports retain full JSON precision.

C1 – cpu-release-cpu-dense-f64-v1-none-prepared-fused1-host-standard

Case

Candidate

A ms

B ms

B/A

95% interval

Verdict

I/M/R

dense-64

native

0.036478500

0.036458500

0.999451732

[0.857970870, 1.176264066]

I

49/51/0

dense-64

numpy-out

0.247816500

0.253927000

1.024657357

[0.953013705, 1.110543751]

I

49/51/0

dense-64

torch-eager-out

0.413731500

0.440459500

1.064602284

[0.934780557, 1.200728111]

I

49/51/0

dense-256

native

0.441359000

0.461859500

1.046448583

[0.957039092, 1.147097782]

I

67/70/1

dense-256

numpy-out

0.754586500

0.799749000

1.059850660

[0.935422520, 1.320781649]

I

67/70/1

dense-256

torch-eager-out

0.902173000

1.030391000

1.142121300

[0.976227476, 1.364493844]

I

67/70/1

dense-1024

native

10.877847500

10.986021000

1.009944385

[0.954194925, 1.062288370]

I

70/73/1

dense-1024

numpy-out

11.584009500

11.688656000

1.009033703

[0.968076061, 1.059169991]

I

70/73/1

dense-1024

torch-eager-out

12.745391000

12.844314000

1.007761472

[0.962674901, 1.062781969]

I

70/73/1

dense-2048

native

74.685592000

76.539489500

1.024822693

[0.971689800, 1.093725354]

I

70/73/1

dense-2048

numpy-out

61.558876500

62.171188000

1.009946762

[0.971609721, 1.100218654]

I

70/73/1

dense-2048

torch-eager-out

65.439674000

66.826310000

1.021189531

[0.985161210, 1.091859318]

I

70/73/1

C2 – cpu-blas-cpu-dense-f64-v1-none-prepared-fused0-host-explicit

Case

Candidate

A ms

B ms

B/A

95% interval

Verdict

I/M/R

dense-64

native

0.060915500

0.060741500

0.997143584

[0.833278434, 1.035042225]

W

49/99/0

dense-64

native-blas

0.046028000

0.044047000

0.956960980

[0.784658332, 1.070160933]

I

50/101/0

dense-64

numpy-out

0.365628500

0.342559500

0.936905903

[0.795059364, 1.067032071]

I

50/101/0

dense-64

torch-eager-out

0.612165500

0.587366500

0.959489713

[0.765387188, 1.213168078]

I

49/99/0

dense-256

native

0.832015000

0.830772000

0.998506037

[0.932858785, 1.051655260]

I

65/131/0

dense-256

native-blas

0.591559000

0.587602000

0.993310895

[0.808875717, 1.095287667]

I

65/131/0

dense-256

numpy-out

1.180804000

1.194225000

1.011365985

[0.914962405, 1.149700282]

I

65/131/0

dense-256

torch-eager-out

1.512979000

1.514156500

1.000778266

[0.839536306, 1.124478410]

I

65/131/0

dense-1024

native

20.453768000

20.201086000

0.987646188

[0.950111738, 1.017758486]

W

69/139/0

dense-1024

native-blas

19.224594500

19.137412500

0.995465080

[0.943557892, 1.039939604]

W

69/139/0

dense-1024

numpy-out

20.989331500

20.834847500

0.992639880

[0.943212008, 1.032535597]

W

69/139/0

dense-1024

torch-eager-out

23.203545500

23.041773000

0.993028113

[0.943142955, 1.034957393]

W

69/139/0

dense-2048

native

148.032970500

151.788187000

1.025367433

[0.925235561, 1.144349838]

I

68/137/0

dense-2048

native-blas

114.550510500

118.164948500

1.031553225

[0.914674573, 1.165967152]

I

68/137/0

dense-2048

numpy-out

117.635083000

121.202122500

1.030322922

[0.891473878, 1.181774149]

I

68/137/0

dense-2048

torch-eager-out

124.830293000

130.597615500

1.046201305

[0.919222680, 1.203120045]

I

68/137/0

C3 – cpu-openmp-cpu-dense-f64-v1-jacobi-prepared-fused1-host-standard

Case

Candidate

A ms

B ms

B/A

95% interval

Verdict

I/M/R

dense-64

native

0.010440500

0.010346500

0.990996600

[0.961697969, 1.020191083]

W

5/7/0

dense-64

numpy-out

0.029012000

0.028550000

0.984075555

[0.960963892, 1.017699023]

W

5/7/0

dense-64

torch-eager-out

0.044832500

0.045174000

1.007617242

[0.969232652, 1.049710303]

W

5/7/0

dense-256

native

0.038686500

0.033943500

0.877399093

[0.706015027, 1.174343343]

I

4/6/0

dense-256

numpy-out

0.063162500

0.058882000

0.932230358

[0.654672503, 1.327264641]

I

4/6/0

dense-256

torch-eager-out

0.090306500

0.082682000

0.915570861

[0.590233346, 1.364876785]

I

4/6/0

dense-1024

native

1.088244000

0.971866000

0.893058910

[0.656242819, 1.295332512]

I

4/6/0

dense-1024

numpy-out

1.397346500

1.452268500

1.039304496

[0.740325484, 1.545161471]

I

4/6/0

dense-1024

torch-eager-out

1.327797500

1.153067500

0.868406139

[0.681905432, 1.154333336]

I

4/6/0

dense-2048

native

7.245700500

7.402156500

1.021592943

[0.951930773, 1.109022282]

I

3/5/0

dense-2048

numpy-out

6.326082500

6.198597000

0.979847639

[0.779014883, 1.067923799]

I

3/5/0

dense-2048

torch-eager-out

7.256162000

7.371341000

1.015873267

[0.952578214, 1.094900050]

I

3/5/0

C4 – cuda-release-cuda-dense-f64-v1-jacobi-api-fused1-host-standard

Case

Candidate

A ms

B ms

B/A

95% interval

Verdict

I/M/R

dense-64

cuda-native

0.973382500

0.970757000

0.997302705

[0.959199197, 1.029853624]

W

5/7/0

dense-64

cuda-cublas

2.599223000

2.552911000

0.982182368

[0.924884842, 1.013225501]

W

5/7/0

dense-64

torch-cuda-eager

2.075354000

2.041694500

0.983781321

[0.942057862, 1.038661441]

W

5/7/0

dense-256

cuda-native

0.871065500

0.876715500

1.006486309

[0.974348428, 1.031659770]

W

4/6/0

dense-256

cuda-cublas

2.494266500

2.505634000

1.004557452

[0.980173710, 1.028987772]

W

4/6/0

dense-256

torch-cuda-eager

1.787447000

1.818148000

1.017175894

[0.973617948, 1.061824775]

I

4/6/0

dense-1024

cuda-native

1.156404000

1.167750000

1.009811450

[0.991715451, 1.028847427]

W

4/6/0

dense-1024

cuda-cublas

2.681247500

2.722697000

1.015459035

[0.988296020, 1.036793228]

W

4/6/0

dense-1024

torch-cuda-eager

1.825714500

1.837918000

1.006684232

[0.968517482, 1.054593297]

I

4/6/0

dense-2048

cuda-native

1.480894500

1.492668000

1.007950262

[0.986748532, 1.027180139]

W

3/5/0

dense-2048

cuda-cublas

2.932352500

2.931416500

0.999680802

[0.980069469, 1.016471723]

W

3/5/0

dense-2048

torch-cuda-eager

2.017374000

2.009443500

0.996068899

[0.951776726, 1.023582800]

W

3/5/0

C5 – cuda-release-cuda-dense-f64-v1-jacobi-api-fused1-device-standard

Case

Candidate

A ms

B ms

B/A

95% interval

Verdict

I/M/R

dense-64

cuda-native

0.700862000

0.688858500

0.982873233

[0.937871677, 1.026712615]

W

5/7/0

dense-64

cuda-cublas

2.334348500

2.336503500

1.000923170

[0.956361980, 1.061953464]

I

5/7/0

dense-64

torch-cuda-eager

2.055879500

2.068202500

1.005994028

[0.948864842, 1.059520854]

I

5/7/0

dense-256

cuda-native

0.672045000

0.676133000

1.006082926

[0.961887822, 1.050206816]

I

4/6/0

dense-256

cuda-cublas

2.272285500

2.299828000

1.012121056

[0.974395476, 1.057859328]

I

4/6/0

dense-256

torch-cuda-eager

1.778866000

1.774317500

0.997443034

[0.956262134, 1.044812401]

W

4/6/0

dense-1024

cuda-native

0.958423000

0.964118000

1.005942053

[0.964407802, 1.054036065]

I

4/6/0

dense-1024

cuda-cublas

2.459574500

2.466921500

1.002987102

[0.967266373, 1.049748515]

W

4/6/0

dense-1024

torch-cuda-eager

1.832807000

1.823488000

0.994915449

[0.953456637, 1.037041800]

W

4/6/0

dense-2048

cuda-native

1.310687000

1.299524500

0.991483474

[0.932765678, 1.051592577]

I

3/5/0

dense-2048

cuda-cublas

2.693607000

2.693731000

1.000046035

[0.940558067, 1.044620375]

W

3/5/0

dense-2048

torch-cuda-eager

1.992150500

1.972030500

0.989900361

[0.950965101, 1.028961607]

W

3/5/0

C6 – cuda-release-cuda-dense-f64-v1-jacobi-session-fused1-host-standard

Case

Candidate

A ms

B ms

B/A

95% interval

Verdict

I/M/R

dense-64

cuda-native

1.055452000

1.051959000

0.996690517

[0.937398389, 1.054995893]

I

5/7/0

dense-64

cuda-cublas

1.090300500

1.067735500

0.979303871

[0.915756349, 1.048680615]

W

5/7/0

dense-64

torch-cuda-eager

2.301979000

2.324570500

1.009813947

[0.944174286, 1.092038733]

I

5/7/0

dense-256

cuda-native

0.962850000

0.947361500

0.983913901

[0.944722994, 1.021418662]

W

4/6/0

dense-256

cuda-cublas

0.952832500

0.957803000

1.005216552

[0.947626180, 1.035667109]

W

4/6/0

dense-256

torch-cuda-eager

2.071974500

2.080366000

1.004050002

[0.943678302, 1.062849749]

I

4/6/0

dense-1024

cuda-native

1.231823500

1.232015500

1.000155866

[0.937151470, 1.059762990]

I

4/6/0

dense-1024

cuda-cublas

1.129131500

1.138191000

1.008023423

[0.952470843, 1.061290913]

I

4/6/0

dense-1024

torch-cuda-eager

2.041687000

2.042009000

1.000157713

[0.918963897, 1.109550688]

I

4/6/0

dense-2048

cuda-native

1.590005500

1.544001000

0.971066452

[0.934676822, 1.017661536]

W

3/5/0

dense-2048

cuda-cublas

1.420603500

1.382065500

0.972872093

[0.949243406, 1.017365711]

W

3/5/0

dense-2048

torch-cuda-eager

2.188302500

2.124503000

0.970845210

[0.924256285, 1.041721268]

W

3/5/0

C7 – cuda-release-cuda-dense-f64-v1-jacobi-prepared-fused1-device-standard

Case

Candidate

A ms

B ms

B/A

95% interval

Verdict

I/M/R

dense-64

cuda-native

0.686690500

0.684183500

0.996349156

[0.900202610, 1.110989711]

I

5/7/0

dense-64

cuda-cublas

0.752553000

0.735382000

0.977183002

[0.892936367, 1.073842090]

I

5/7/0

dense-64

torch-cuda-eager

2.075066000

2.012004000

0.969609641

[0.862456567, 1.084288885]

I

5/7/0

dense-256

cuda-native

0.635981000

0.644305000

1.013088441

[0.961525500, 1.074139747]

I

4/6/0

dense-256

cuda-cublas

0.653569500

0.652732500

0.998719340

[0.934631061, 1.068020189]

I

4/6/0

dense-256

torch-cuda-eager

1.712591500

1.712064000

0.999691987

[0.942103080, 1.081600347]

I

4/6/0

dense-1024

cuda-native

0.929282000

0.920169000

0.990193504

[0.878957070, 1.034421628]

W

4/6/0

dense-1024

cuda-cublas

0.809870000

0.795822000

0.982654006

[0.858664074, 1.037646781]

W

4/6/0

dense-1024

torch-cuda-eager

1.799256500

1.710087500

0.950441196

[0.785608551, 1.075179618]

I

4/6/0

dense-2048

cuda-native

1.320069500

1.328411500

1.006319364

[0.982908623, 1.042312209]

W

3/5/0

dense-2048

cuda-cublas

1.151646000

1.149937500

0.998516471

[0.971460097, 1.056480319]

I

3/5/0

dense-2048

torch-cuda-eager

1.843685500

1.878265500

1.018755910

[0.950309287, 1.108193191]

I

3/5/0

Eight process medians per side

The eight values in each cell are process-block medians in milliseconds, ordered exactly as the comparator’s baseline_files or current_files array. For either side the first four belong to window 1 and the last four to window 2; within a window they retain actual ABBA then BAAB slot order. These are descriptive process observations, not independent whole-host replications or a replacement estimator for the primary decision.

C1 process observations

Case

Candidate

A1–A8 ms

B1–B8 ms

dense-64

native

0.034863000, 0.035210500, 0.039969500, 0.033544500, 0.036896000, 0.028922500, 0.040526500, 0.044745500

0.039333000, 0.032091500, 0.035785000, 0.037543500, 0.035688000, 0.030919500, 0.037484500, 0.036767000

dense-64

numpy-out

0.246572500, 0.243467500, 0.268640500, 0.238535500, 0.257019000, 0.200451500, 0.272417000, 0.297465500

0.321841500, 0.244299000, 0.241008000, 0.257795000, 0.269883000, 0.246159000, 0.258627000, 0.256960000

dense-64

torch-eager-out

0.404574000, 0.416888000, 0.454255000, 0.360999500, 0.432832000, 0.305040000, 0.535784500, 0.578903000

0.556850500, 0.416606500, 0.410318000, 0.436422000, 0.446673000, 0.425699000, 0.450299500, 0.435097500

dense-256

native

0.450488000, 0.437183500, 0.463028000, 0.418411000, 0.458039000, 0.379728000, 0.560662500, 0.531689000

0.556308500, 0.445460500, 0.478523500, 0.463751500, 0.475161000, 0.434665000, 0.453568000, 0.445390500

dense-256

numpy-out

0.756385000, 0.766068000, 0.840736500, 0.627218000, 0.876137000, 0.519323500, 0.880085500, 0.910001000

0.796256500, 0.786690500, 0.721878500, 0.865352500, 0.717317000, 0.787590000, 0.850203000, 0.825656500

dense-256

torch-eager-out

0.903430000, 0.847167000, 1.076970500, 0.808805000, 0.994489000, 0.644816000, 1.168479500, 1.274132500

1.253483500, 0.959984500, 1.005450000, 1.010615500, 1.024958500, 1.044639000, 0.996766500, 1.039443000

dense-1024

native

10.611854500, 11.900283500, 10.535237000, 10.662038000, 10.778511500, 10.864079500, 11.459580000, 10.452886000

11.983497000, 10.901240500, 11.343744500, 11.577150500, 11.296724500, 10.245993500, 10.308624500, 10.315050500

dense-1024

numpy-out

11.738073000, 12.361164500, 11.576706000, 11.208671500, 11.396704000, 11.080290000, 11.707361000, 11.391166000

12.590694500, 11.440718000, 11.789707000, 12.494155000, 12.461959500, 11.290409500, 11.105924000, 11.168976500

dense-1024

torch-eager-out

13.094442000, 13.721380500, 12.449621500, 12.031225500, 12.802075000, 12.399951000, 13.401078500, 12.439695000

14.039811500, 13.647669500, 12.876563000, 13.522929000, 13.116406500, 12.241545000, 12.194187500, 12.220482500

dense-2048

native

78.086862000, 72.193961500, 77.527322000, 75.435834000, 74.000481000, 74.495811500, 74.943425500, 70.459720500

96.208170000, 77.903885500, 81.018674000, 73.159933500, 72.095060000, 76.715016500, 71.795599000, 80.053109500

dense-2048

numpy-out

66.081534000, 60.454254000, 64.414973000, 59.438171000, 62.246008500, 59.620978000, 62.627675000, 61.454071000

78.940696500, 63.606289500, 60.289337500, 61.941228000, 60.292884500, 61.888964500, 60.662190500, 62.981402500

dense-2048

torch-eager-out

68.309388500, 69.527296500, 71.377085000, 64.139753000, 66.375117500, 65.565928500, 63.925974000, 64.815065000

91.092626500, 68.107557500, 67.788995500, 65.213681000, 64.954951000, 67.455641500, 65.763718500, 69.265656500

C2 process observations

Case

Candidate

A1–A8 ms

B1–B8 ms

dense-64

native

0.052457500, 0.060795500, 0.062797000, 0.064647000, 0.059932000, 0.059742000, 0.062656500, 0.062066000

0.062592000, 0.048838000, 0.062016500, 0.061497000, 0.062031000, 0.061875000, 0.049893000, 0.063677000

dense-64

native-blas

0.040221500, 0.045990500, 0.041195500, 0.049693000, 0.045913000, 0.045085500, 0.048376500, 0.046600500

0.047200000, 0.035205000, 0.046908000, 0.048127000, 0.054950000, 0.047343500, 0.034457500, 0.039453000

dense-64

numpy-out

0.303790000, 0.347884500, 0.341876500, 0.389832500, 0.373756000, 0.337599500, 0.387809500, 0.378109500

0.365808500, 0.270206500, 0.388960000, 0.378348000, 0.380363000, 0.376568500, 0.297512500, 0.350238000

dense-64

torch-eager-out

0.523617000, 0.577233000, 0.675729500, 0.652441000, 0.641911500, 0.513616000, 0.697876000, 0.632148500

0.665561000, 0.406075000, 0.689107000, 0.671817500, 0.693466500, 0.714361000, 0.501410000, 0.559348500

dense-256

native

0.805921000, 0.835670000, 0.830871000, 0.895662000, 0.812581500, 0.816715000, 0.850265500, 0.851611000

0.844728000, 0.730272000, 0.830772000, 0.850372500, 0.878237000, 0.832477500, 0.774864000, 0.927911500

dense-256

native-blas

0.535125000, 0.748705000, 0.547132500, 0.591166500, 0.583421500, 0.549045500, 0.575121500, 0.806010000

0.586864000, 0.478805500, 0.525086500, 0.755438000, 0.601404500, 0.590517500, 0.525746000, 0.630983000

dense-256

numpy-out

1.143487500, 1.105344500, 1.192008500, 1.377552000, 1.220759500, 1.181670500, 1.229261500, 1.087061000

1.292224500, 1.009186500, 1.201823500, 1.107744500, 1.287730000, 1.305933000, 1.089487500, 1.447479500

dense-256

torch-eager-out

1.347873500, 1.648238500, 1.739750500, 1.542157000, 1.533026000, 1.420811000, 1.613581000, 1.579182000

1.622044500, 1.041704000, 1.565463000, 1.592122500, 1.696052500, 1.405526000, 1.273283000, 1.775769500

dense-1024

native

20.010429500, 20.277092000, 21.126848500, 23.511977500, 20.714939000, 20.429846000, 20.323844000, 19.476441000

20.486658500, 20.727956500, 20.144060500, 21.555562000, 20.475919000, 18.777469500, 19.599590000, 19.716078500

dense-1024

native-blas

19.046026000, 18.702500000, 20.454149500, 21.555899000, 19.457936000, 19.531430000, 18.864539000, 17.614002000

19.597870500, 19.683844500, 19.335895000, 19.941817500, 19.542984000, 17.910844500, 18.502737500, 19.077455000

dense-1024

numpy-out

20.226633500, 20.532795500, 21.753460500, 23.860933000, 21.688469500, 20.777657000, 20.185352000, 20.503456500

21.005725500, 21.223596500, 20.448095500, 21.894136000, 22.023093000, 19.681422500, 19.982001000, 20.632702500

dense-1024

torch-eager-out

21.795185000, 23.166457500, 23.982652500, 26.824400000, 24.039930500, 23.268672500, 22.568099500, 21.965041000

23.274487000, 23.199040000, 23.030598000, 24.684153500, 23.427295000, 21.375570000, 22.319379000, 23.197778000

dense-2048

native

148.001566000, 147.432839500, 139.506271500, 142.186951000, 185.588234000, 176.226529500, 143.792315000, 137.086221500

143.142068000, 146.659050500, 143.030876000, 152.773396500, 161.842878500, 212.869596500, 147.181861000, 140.529140500

dense-2048

native-blas

111.019615500, 113.300648000, 113.790834500, 111.775844000, 143.583941000, 137.015266000, 117.297930000, 108.982142000

112.228563500, 116.887367500, 112.403745500, 122.703934500, 134.110912500, 174.987923000, 108.170463000, 106.487269500

dense-2048

numpy-out

114.614108000, 116.103846500, 115.265909500, 115.941028000, 146.726132500, 145.748993000, 119.685960500, 114.987416500

111.229478000, 122.023513500, 115.026734500, 121.334125500, 134.574519500, 183.147272000, 116.313342000, 114.874081000

dense-2048

torch-eager-out

120.556644500, 123.634555500, 122.694030000, 121.781264500, 168.438647500, 150.415628500, 119.204039000, 121.784406500

124.029810500, 131.368845000, 124.004652000, 127.956085000, 149.462738000, 186.075534500, 126.793812500, 122.240494000

C3 process observations

Case

Candidate

A1–A8 ms

B1–B8 ms

dense-64

native

0.011725000, 0.010162000, 0.010372000, 0.010030500, 0.010595000, 0.010435000, 0.010422500, 0.010442000

0.010549500, 0.010417000, 0.010401500, 0.010244000, 0.010181500, 0.010080500, 0.010293000, 0.011429500

dense-64

numpy-out

0.032217500, 0.028953500, 0.028563000, 0.027921500, 0.028783500, 0.028607500, 0.029028500, 0.029040000

0.028976000, 0.029176000, 0.028516500, 0.028280000, 0.028037500, 0.028477500, 0.028174000, 0.030966500

dense-64

torch-eager-out

0.049517500, 0.044644000, 0.044625000, 0.043959000, 0.044992500, 0.043547000, 0.043965000, 0.045521000

0.043990500, 0.046891500, 0.044528000, 0.044772000, 0.043387000, 0.044567000, 0.044626000, 0.049738000

dense-256

native

0.032212500, 0.030298500, 0.639454500, 0.045252500, 0.093240500, 0.040354000, 0.032674500, 0.037433000

0.045586500, 0.031061500, 0.034607500, 0.039941000, 0.030278000, 0.043906000, 0.028943500, 0.030499500

dense-256

numpy-out

0.051136500, 0.046824000, 0.143120000, 0.080212500, 0.163523500, 0.068906500, 0.062035000, 0.056071500

0.091220500, 0.052049000, 0.071950500, 0.044645500, 0.054875500, 0.089058500, 0.051149500, 0.058215500

dense-256

torch-eager-out

0.066494500, 0.061189500, 0.704291000, 0.112444000, 0.212067500, 0.110353500, 0.072216000, 0.071343000

0.116433500, 0.078399000, 0.089643500, 0.077447000, 0.060044500, 0.117491500, 0.078203000, 0.064983000

dense-1024

native

1.341747500, 1.256959000, 1.194963000, 0.710565500, 1.210974500, 0.727304000, 0.778420500, 1.165052500

0.761817000, 0.738948000, 0.825430500, 0.997547500, 1.249928000, 0.749993500, 0.778445000, 1.266317500

dense-1024

numpy-out

1.276987500, 1.273802500, 1.187841000, 2.191399000, 2.082795500, 2.036614500, 2.067059500, 1.229934000

2.016046500, 1.889496500, 2.090487500, 1.250801500, 1.200829500, 2.255819500, 2.051360500, 1.267919000

dense-1024

torch-eager-out

1.457255500, 1.376946500, 1.483178000, 0.893649500, 1.530526500, 0.927867000, 0.966329500, 1.385263500

0.892529000, 0.872634500, 1.162492500, 1.124062500, 1.342858500, 0.859049000, 0.870773500, 1.518936500

dense-2048

native

7.422519500, 8.150682500, 6.335358500, 7.829901000, 6.925826500, 7.003889500, 7.375794000, 7.676340000

7.311218000, 7.079666000, 7.936069000, 7.026778500, 6.829259000, 7.673926000, 7.204788500, 8.351912000

dense-2048

numpy-out

6.641674500, 6.374474000, 6.076810000, 5.328452000, 6.085180500, 6.331504500, 6.132924500, 7.062382500

4.927955500, 4.568638500, 6.580330000, 4.885934000, 6.198597000, 6.471636000, 6.134253000, 7.321415500

dense-2048

torch-eager-out

7.085855000, 7.086835000, 6.318175500, 7.585919500, 7.461632000, 7.354397500, 7.555504500, 8.034053500

7.181113500, 7.278538500, 7.866048000, 7.476059500, 6.954293500, 7.358854000, 7.291377000, 8.553031500

C4 process observations

Case

Candidate

A1–A8 ms

B1–B8 ms

dense-64

cuda-native

0.976143000, 0.949691500, 0.952035000, 0.991161000, 0.998563000, 0.982062500, 0.975260000, 0.985992000

0.876854500, 0.974678500, 0.905259000, 0.946858500, 1.007061000, 1.008706500, 1.028120500, 0.970945500

dense-64

cuda-cublas

2.607696000, 2.542307000, 2.557175500, 2.578921000, 2.664830000, 2.744041000, 2.552429000, 2.642349500

2.344620000, 2.499267000, 2.328487000, 2.403569000, 2.613824500, 2.647318000, 2.622622500, 2.623077500

dense-64

torch-cuda-eager

1.980618500, 1.969853500, 1.979269500, 1.959567000, 2.113384000, 2.142670000, 2.081057000, 2.145244000

1.932675500, 2.060033500, 1.926367000, 1.998845500, 2.079747000, 2.132832500, 2.027840500, 2.125937500

dense-256

cuda-native

0.851905500, 0.853903500, 0.862256000, 0.866259500, 0.876144000, 0.889931500, 0.897293500, 0.911229500

0.855503000, 0.880720500, 0.860502500, 0.872413000, 0.889819000, 0.874563000, 0.915609500, 0.872218000

dense-256

cuda-cublas

2.456307500, 2.448579000, 2.480162500, 2.478660500, 2.467317000, 2.559714500, 2.558647000, 2.580137500

2.415924500, 2.523289000, 2.457862000, 2.461346000, 2.535822500, 2.549416500, 2.566963000, 2.562143500

dense-256

torch-cuda-eager

1.676317000, 1.726075500, 1.734398500, 1.742137500, 1.793414500, 1.891293500, 1.786850000, 1.900978500

1.724811000, 1.864139500, 1.739151500, 1.767076500, 1.953556000, 1.812269000, 1.879501000, 1.824911000

dense-1024

cuda-native

1.132669000, 1.132016500, 1.125673500, 1.158574500, 1.191535500, 1.191804500, 1.159253000, 1.167353000

1.142580500, 1.166621500, 1.143958000, 1.169321000, 1.156803500, 1.199830500, 1.191224500, 1.189778000

dense-1024

cuda-cublas

2.641191000, 2.601588000, 2.649325500, 2.708300000, 2.696074000, 2.777890500, 2.675384000, 2.722685500

2.660499000, 2.755576500, 2.602457000, 2.669799000, 2.780062000, 2.799525500, 2.739350000, 2.765297000

dense-1024

torch-cuda-eager

1.731847000, 1.743886000, 1.758826500, 1.891018000, 1.846597500, 1.889727000, 1.808225000, 1.941500500

1.819836000, 1.883837000, 1.763989500, 1.772376500, 1.786962000, 1.950385500, 1.826557000, 1.903889000

dense-2048

cuda-native

1.470833000, 1.469142500, 1.502085000, 1.479514500, 1.474236500, 1.507788000, 1.482153000, 1.489214000

1.472334000, 1.492035500, 1.462012500, 1.461981500, 1.488686500, 1.506625000, 1.546110000, 1.523580500

dense-2048

cuda-cublas

2.922558000, 2.900387500, 2.908624000, 2.908229500, 2.942451000, 2.974046000, 2.896662000, 3.036700000

2.854246000, 2.920671000, 2.869858000, 2.881927500, 2.913216000, 3.017363000, 2.987138500, 2.978406500

dense-2048

torch-cuda-eager

1.983146500, 1.962876500, 1.948925500, 1.932407000, 2.002816000, 2.099613000, 2.104754000, 2.161927000

1.991019500, 2.045032000, 2.007544000, 1.988578000, 1.979927000, 2.005490000, 2.011573000, 2.064135500

C5 process observations

Case

Candidate

A1–A8 ms

B1–B8 ms

dense-64

cuda-native

0.756791000, 0.691030000, 0.657041500, 0.655092000, 0.724902500, 0.658757000, 0.706506000, 0.706460000

0.710017000, 0.701116000, 0.675827500, 0.636075000, 0.698282500, 0.665916000, 0.668224000, 0.720749000

dense-64

cuda-cublas

2.511867000, 2.318306500, 2.160590000, 2.078786500, 2.366746000, 2.040850500, 2.391917000, 2.379148500

2.409987500, 2.379011500, 2.317249000, 2.055190000, 2.345526500, 2.172257500, 2.328450500, 2.376427500

dense-64

torch-cuda-eager

2.174865000, 1.959278000, 1.975350000, 1.987072000, 2.067515000, 1.993320000, 2.144297000, 2.066993500

2.229091000, 2.018002000, 1.977033500, 1.860167500, 2.068802000, 2.060211500, 2.006243000, 2.187362500

dense-256

cuda-native

0.718240500, 0.650980000, 0.654710500, 0.650864500, 0.693693500, 0.675327500, 0.667311000, 0.669096500

0.728553500, 0.663016500, 0.652378500, 0.664527500, 0.663932000, 0.682754500, 0.679008000, 0.679376500

dense-256

cuda-cublas

2.470525500, 2.242598500, 2.220877000, 2.206293000, 2.285800000, 2.272715000, 2.285420500, 2.301141500

2.481290000, 2.286401500, 2.245318500, 2.204976000, 2.281471000, 2.357268500, 2.277582500, 2.382239000

dense-256

torch-cuda-eager

1.941165500, 1.742722500, 1.730577500, 1.701637500, 1.835770500, 1.776708500, 1.761172000, 1.768981000

1.928669500, 1.716596000, 1.658214000, 1.714363500, 1.777292000, 1.839180000, 1.773860000, 1.782791000

dense-1024

cuda-native

1.020732500, 0.923289500, 0.933372000, 0.952477000, 0.958310500, 0.987663500, 0.962260500, 0.954437500

1.031538000, 0.942935000, 0.926795500, 0.931413000, 0.967317500, 0.956157000, 0.957643500, 0.997703500

dense-1024

cuda-cublas

2.623421000, 2.406624500, 2.355571000, 2.428770000, 2.420221500, 2.518552500, 2.460275500, 2.475390500

2.665447000, 2.403424000, 2.395508500, 2.385966500, 2.464929500, 2.502450000, 2.469516000, 2.550288000

dense-1024

torch-cuda-eager

2.025963000, 1.753304000, 1.768614500, 1.768458500, 1.825223500, 1.892802000, 1.850138000, 1.836816000

1.945506500, 1.765888000, 1.785111000, 1.745575500, 1.844140000, 1.845400500, 1.796078500, 1.862946000

dense-2048

cuda-native

1.459068000, 1.239852000, 1.251723000, 1.249078000, 1.352463000, 1.342723000, 1.350989000, 1.252674500

1.259080500, 1.244780500, 1.231785000, 1.313897500, 1.343711500, 1.371732000, 1.293001500, 1.304332500

dense-2048

cuda-cublas

2.978474000, 2.554766500, 2.628190500, 2.578413000, 2.773440500, 2.776611000, 2.809223500, 2.651693000

2.630277000, 2.606518000, 2.534772500, 2.694192000, 2.779888500, 2.853963000, 2.716538500, 2.687824000

dense-2048

torch-cuda-eager

2.117021500, 1.863862500, 1.935893000, 1.955716500, 1.985149500, 2.000111500, 2.045383500, 1.932164000

1.939356500, 1.901047000, 1.888920500, 1.950274000, 1.998780500, 2.080354500, 2.001809000, 1.999463500

C6 process observations

Case

Candidate

A1–A8 ms

B1–B8 ms

dense-64

cuda-native

1.058852000, 0.960998000, 1.192425500, 1.021661000, 1.089756000, 1.058205000, 1.005072000, 1.075331000

1.017685000, 1.015317000, 1.097058000, 1.048227500, 1.018771000, 1.160319000, 1.038598500, 1.116000500

dense-64

cuda-cublas

1.112173000, 1.015508000, 1.248267500, 1.048002500, 1.099931000, 1.120944500, 1.035932000, 1.125719500

1.033318500, 1.001883500, 1.081862500, 1.069224000, 1.037975000, 1.195224000, 1.082638500, 1.151985000

dense-64

torch-cuda-eager

2.396633000, 2.240985500, 2.767619500, 2.189033000, 2.370450000, 2.291004500, 2.188399500, 2.362297000

2.252341500, 2.228690500, 2.538618000, 2.263202500, 2.232969000, 2.596406000, 2.304693500, 2.629142000

dense-256

cuda-native

0.960969000, 0.947113500, 1.056390000, 1.024687000, 0.946758500, 0.954209000, 0.925633500, 0.940284000

0.912178000, 0.920938000, 0.950869500, 0.925538000, 0.940972500, 0.987976500, 0.946692000, 0.999968500

dense-256

cuda-cublas

0.948960000, 0.926711500, 1.059967500, 0.993359500, 0.942748500, 0.956602500, 0.924325000, 0.931454500

0.911110000, 0.913101500, 0.949858500, 0.950523000, 0.956357500, 1.000899000, 0.966539500, 0.987141000

dense-256

torch-cuda-eager

2.101947500, 2.020676000, 2.348385000, 2.124956000, 2.056069500, 1.973004000, 2.030604500, 2.032169500

1.961052000, 1.949164500, 2.149007000, 1.996247000, 2.018325500, 2.269686500, 2.091043500, 2.287014000

dense-1024

cuda-native

1.233264500, 1.171732000, 1.387385500, 1.307418000, 1.143720500, 1.256560000, 1.214936000, 1.189288000

1.146746000, 1.195410000, 1.353063000, 1.208310500, 1.247749000, 1.312411500, 1.233504500, 1.296120500

dense-1024

cuda-cublas

1.129131500, 1.073292000, 1.223961000, 1.189760000, 1.050132500, 1.152245500, 1.126715500, 1.101087500

1.060200000, 1.094270000, 1.220965500, 1.095484000, 1.127544500, 1.187061000, 1.167152000, 1.200936000

dense-1024

torch-cuda-eager

2.154927500, 1.938299000, 2.389294500, 2.188384500, 1.748322500, 2.068449500, 2.003670500, 1.957254500

1.776475000, 1.950867500, 2.277662000, 1.968101500, 1.985837000, 2.370008000, 2.013941000, 2.212728500

dense-2048

cuda-native

1.566349000, 1.545473500, 1.647134000, 1.674490500, 1.566443500, 1.558577500, 1.570602000, 1.603178000

1.506743500, 1.478755000, 1.718992000, 1.516905500, 1.567551500, 1.613867000, 1.514526000, 1.547271000

dense-2048

cuda-cublas

1.406331500, 1.357600000, 1.463036500, 1.460465000, 1.391031000, 1.410578000, 1.406675500, 1.397940500

1.368574000, 1.320223000, 1.513920500, 1.352095000, 1.416017500, 1.439498500, 1.380636500, 1.364791000

dense-2048

torch-cuda-eager

2.253799000, 2.152138000, 2.214669000, 2.285615500, 2.138135000, 2.150512000, 2.135652000, 2.187812000

2.094800500, 2.017490500, 2.435348000, 2.077767000, 2.175602000, 2.307237000, 2.067454500, 2.037637500

C7 process observations

Case

Candidate

A1–A8 ms

B1–B8 ms

dense-64

cuda-native

0.602499000, 0.676204500, 0.641782500, 0.840580500, 0.728921500, 0.754733000, 0.702643000, 0.796951000

0.678258000, 0.648314500, 0.703109000, 0.894646000, 0.638390000, 0.673202500, 0.860620500, 0.658208000

dense-64

cuda-cublas

0.655174500, 0.736147500, 0.728911500, 1.026963500, 0.755851500, 0.756554500, 0.769478000, 0.973403000

0.714897500, 0.716759500, 0.759471500, 0.965431500, 0.687136500, 0.719517000, 0.932687500, 0.718472000

dense-64

torch-cuda-eager

1.809247500, 2.190159000, 1.956555000, 2.783302500, 2.089637000, 2.196478500, 1.949247500, 2.543216000

1.943441000, 1.917333500, 1.966230000, 2.761847500, 1.999103500, 2.027765500, 2.721109000, 2.005371000

dense-256

cuda-native

0.596219500, 0.627107000, 0.637750500, 0.783451500, 0.643968000, 0.616033000, 0.762353000, 0.640074500

0.664694000, 0.636190500, 0.630275000, 0.855263000, 0.629816000, 0.621216500, 0.679869500, 0.635617000

dense-256

cuda-cublas

0.611068500, 0.686863000, 0.648642500, 0.715774000, 0.635426000, 0.637167500, 0.857536500, 0.682934000

0.667691000, 0.645010000, 0.671346000, 0.910342500, 0.621762500, 0.628735500, 0.696579000, 0.631433500

dense-256

torch-cuda-eager

1.570924500, 1.749644500, 1.721078500, 1.992389000, 1.665163500, 1.698908500, 1.900321500, 1.808255500

1.740680500, 1.626315500, 1.706004000, 2.377185000, 1.692031000, 1.671273500, 1.923590000, 1.651795500

dense-1024

cuda-native

0.882587500, 0.912479000, 0.893514500, 1.120369000, 0.948060000, 1.051670500, 0.915294000, 1.036278500

0.948602500, 0.899119000, 0.896024000, 0.943596000, 0.991591000, 0.911206000, 0.961334500, 0.908214500

dense-1024

cuda-cublas

0.757975000, 0.795062500, 0.769881500, 1.017732500, 0.829684000, 0.930255500, 0.796767000, 0.934736500

0.817977500, 0.773250000, 0.789941500, 0.793373000, 0.947414000, 0.779345500, 0.831898000, 0.786994500

dense-1024

torch-cuda-eager

1.548684500, 1.790191000, 1.614466500, 2.452524000, 1.706811500, 2.190169500, 1.696837000, 2.057936000

1.756196500, 1.636901000, 1.684351000, 1.672165000, 2.179268500, 1.710087500, 1.895507500, 1.631982500

dense-2048

cuda-native

1.266488000, 1.334597500, 1.298695000, 1.425639500, 1.332806000, 1.322630500, 1.302552000, 1.315012000

1.343006500, 1.265816500, 1.301113500, 1.460985000, 1.382675000, 1.367156000, 1.318720000, 1.313486500

dense-2048

cuda-cublas

1.098609000, 1.177786500, 1.140402500, 1.235242500, 1.172779500, 1.153630500, 1.133032500, 1.162116000

1.161543000, 1.091221500, 1.124997000, 1.243638000, 1.265490500, 1.212367000, 1.139346500, 1.131442500

dense-2048

torch-cuda-eager

1.745136000, 1.959729500, 1.794697000, 2.179952000, 1.804593000, 1.964306000, 1.790214000, 1.834134500

1.875910500, 1.622160000, 1.804604500, 2.242824500, 2.171582000, 2.153555500, 1.900842500, 1.800214000

Reproduce the declared work

The public make pyprof commands below reproduce the seven declared input/mode/scope configurations through the same fixtures, adapters, true-residual checks and Williams scheduler exercised by pytest. Each command is one artifact; it does not by itself reproduce the balanced confidence calibration. Use separate build directories and retain each output under a fresh name. Build before the quiet timing campaign; make pyprof has a build prerequisite and must not target an immutable stage as BUILD_DIR. Exact historical reproduction also requires the solver/tool revisions, compiler/runtime/provider controls and frozen identities above.

common='--seed 20261017 --warmups 3 --rtol 1e-8 --atol 1e-12 --maxiter 2000 --residual-refresh 50'
cpu_refs='--implementations native,numpy-out,torch-eager-out'
blas_refs='--implementations native,native-blas,numpy-out,torch-eager-out'
taskset -c 0 make pyprof PRESET=cpu-release PYTHON=.venv/bin/python \
    BUILD_DIR=build/reproduce-confidence/cpu-release \
    CMAKE_ARGS=-DCMAKE_CXX_COMPILER=/usr/bin/c++ \
    PYPROF_SUITE=cpu-dense-f64-v1 PYPROF_SCOPE=prepared PYPROF_THREADS=1 \
    PYPROF_ROUNDS=36 PYPROF_REPETITIONS=1 \
    PYPROF_ARGS="$common $cpu_refs --algorithm standard --preconditioner none --fused"
taskset -c 0 make pyprof PRESET=cpu-blas PYTHON=.venv/bin/python \
    BUILD_DIR=build/reproduce-confidence/cpu-blas \
    CMAKE_ARGS=-DCMAKE_CXX_COMPILER=/usr/bin/c++ \
    PYPROF_SUITE=cpu-dense-f64-v1 PYPROF_SCOPE=prepared PYPROF_THREADS=1 \
    PYPROF_ROUNDS=36 PYPROF_REPETITIONS=1 \
    PYPROF_ARGS="$common $blas_refs --algorithm explicit --preconditioner none --no-fused"
taskset -c 0,2 make pyprof PRESET=cpu-openmp PYTHON=.venv/bin/python \
    BUILD_DIR=build/reproduce-confidence/cpu-openmp \
    CMAKE_ARGS=-DCMAKE_CXX_COMPILER=/usr/bin/c++ \
    PYPROF_SUITE=cpu-dense-f64-v1 PYPROF_SCOPE=prepared PYPROF_THREADS=2 \
    PYPROF_ROUNDS=36 PYPROF_REPETITIONS=1 \
    PYPROF_ARGS="$common $cpu_refs --algorithm standard --preconditioner jacobi --fused"
for selection in api-host api-device session-host prepared-device; do
    scope=${selection%-*}
    mode=${selection#*-}
    env -u CUDA_LAUNCH_BLOCKING CUBLAS_WORKSPACE_CONFIG=:4096:8 \
        taskset -c 0 make pyprof PRESET=cuda-release PYTHON=.venv-cuda/bin/python \
        BUILD_DIR=build/reproduce-confidence/cuda-release \
        CUDA_ROOT="$PWD/.deps/cuda-12.8" CUDA_HOST_COMPILER=/usr/bin/g++-13 \
        CUDA_ARCHITECTURES=75 CMAKE_ARGS=-DCMAKE_CXX_COMPILER=/usr/bin/g++-13 \
        PYPROF_SUITE=cuda-dense-f64-v1 PYPROF_SCOPE="$scope" PYPROF_THREADS=1 \
        PYPROF_ROUNDS=36 PYPROF_REPETITIONS=1 \
        PYPROF_OUTPUT="build/reproduce-confidence/cuda-$selection.json" \
        PYPROF_ARGS="$common --implementations cuda-native,cuda-cublas,torch-cuda-eager --device cuda --cuda-device 0 --cuda-execution-mode $mode --algorithm standard --preconditioner jacobi --fused"
done

For the full balanced procedure, follow Balanced process campaigns: create one input plan per preset, declaring both aliases of the same frozen stage and the configurations above. Freeze the plan with --plan-only, then execute it once in a new directory. The original four input plans and expanded schedules remain in the evidence directory below. Do not execute an already started output directory or replace it after seeing its outcome. A repeat is a new, separately declared campaign whose outcome must remain alongside this one.

Evidence identities

All paths in this section are relative to the retained local evidence root build/reliable-performance-20261018/. Manifests bind every raw artifact and log digest, telemetry snapshot, launch command and comparison report; plans bind complete stage payloads, compiled-source declarations, interpreter hashes and tool hashes. They preserve the original successful execution’s provenance. This report does not rerun timing or recompute confidence intervals.

File

SHA256

calibration-analysis.json

4974aaeb7bd314a2c39ec27b9bce81d940b0b246c9540442981a09a5cd0f9ef7

source-baseline.json

09403d4df478aec01c5aa72fe4893b5ee43191da353eb80c1619aab1f034d218

tool-source.json

63f9e0a3d658bf83a16887881737dfaeecdec0864fe4c9a1a029b42acc43ec0d

calibration-execution-summary.json

ed7def359c6a05e000e8c3cf92ba2b7030f74fdf32f7c263ba376760da8013f4

additional-observer-query.json

c1bd5c48383c9a9598e2b75b294f3d17c6063d9064eb3b43186affa262669152

setup-summary.json

9ee0ed4a37181d3af5e247fdb6700068b68d7bbd2069c2ff91d5231a04be214f

setup-cpu-release.log

2667d6d5bd91e4a832752b04cac8075daef692785c06b2b3495484ab3bf287f9

setup-fixed-summary.json

d03db540537c9b7953d1bacef0df7cc241a82c7e5dd5f89d3a76d34239e7ab28

setup-fixed-cpu-release.log

d88b60fb9d192470087cf14b036ef4772f9a0b2a08e7a4811c4c1ea0cd598183

setup-fixed-cpu-blas.log

a4645f8bf0a22e66bbb861ffd4b3c7523ed97fc0a6c2fef57a2e1689e9feb321

setup-fixed-cpu-openmp.log

fe13e0b32153e312652d0d790be06ffa5ea7c7c95801319d374c41e0b0187325

setup-fixed-cuda-release.log

2683cb572a8ab3056023721c519ef7bc5ad99675b1da6c4fabdc10913a539ac4

setup_baseline.py

4cf3f8e44889ec6e6a2d3050b58917b274b2289c43ecb0c9c0be2b911c63185c

setup_baseline_fixed.py

9f5b2a40201eea95c92e2397ed5d59f75918c32c669cfc5d5b730d9fb7e634ff

run_calibrations.py

11c01c4ddb5b20084cf619f502ccdc066bd013e633e1a846c254843cb7632170

freeze_stage.py

6e4e06991f095c31d073df67ac2418d2eee52acc4d1226b4e47f108c1d20a2b4

protocol.py

e9d8ee8e95b06c5a41416f807615b5f23020775702f4cb838dce7dd687ebcb0a

cpu-release campaign and stage

File

SHA256

inputs/calibration-cpu-release.json

f150533756be7710643961ff54919454358bbc4557f4927f9f5870188e08a209

calibration-cpu-release/plan.json

9f78ef6014d36a7d797c3f425ce50230824ecdc7f7ce02611daaddb28643da18

calibration-cpu-release/manifest.json

9635ade929760cd3787901edfd097782f5fbc92bc3e9beef490f87ea7be76ab9

calibration-cpu-release/reports/cpu-release-cpu-dense-f64-v1-none-prepared-fused1-host-standard.json

8cc4f174702bed6d2b3e5896c8aa7e5fd211456f44cb8af64e0abbf76d134d36

baseline-stages/cpu-release/source-files.json

de44d54de1c7133d20fbfb3099a8e412d489a4f5e9f59fb1e53e1d8fee2d2cf8

baseline-stages/cpu-release/libcg_core.a

e47471ca51084098a11eee8f0b45bb1b06b8a78b27f0b868408d8e709de6f3da

baseline-stages/cpu-release/compile_commands.json

eb4fd1947170009fe5cc972f3f792e928755bab5bb342b0594a32ebc3eeee632

baseline-stages/cpu-release/identity.json

65d54c622017689d2eabdd543d5cbdc15e70b2030088512972d9e82248883f38

baseline-stages/cpu-release/CMakeCache.txt

3c3a873efe97716f8670ca14a727c4dec59b030cfd8e2d2cbc3d12b9f298fcf7

baseline-stages/cpu-release/python/cgpy/_core.cpython-312-x86_64-linux-gnu.so

76168203038c854eec161fb28eac73b832ec0d0341d76e0266dfdc3514318611

cpu-blas campaign and stage

File

SHA256

inputs/calibration-cpu-blas.json

aac26ae4643fd417a1674e16302a1497c52097fe75ebb69efa770535382754ae

calibration-cpu-blas/plan.json

9cc77c414382eb4b27bb983ff3535ffea46c73081a7fad1ab7a3dd1d23f7ce27

calibration-cpu-blas/manifest.json

5c8f114588219f50c619bb6619c12be28257e4909c0b775818cbf1f1eff849bf

calibration-cpu-blas/reports/cpu-blas-cpu-dense-f64-v1-none-prepared-fused0-host-explicit.json

ba7f317df061a084b408e9aa99479d25118f636879051ade6fb1058bbe0da2e5

baseline-stages/cpu-blas/source-files.json

de44d54de1c7133d20fbfb3099a8e412d489a4f5e9f59fb1e53e1d8fee2d2cf8

baseline-stages/cpu-blas/libcg_core.a

fe9159eb82deff9837978f28fbd51b57ad0aeb097bf68fecb3995e60d4ce5352

baseline-stages/cpu-blas/compile_commands.json

7cd5720a683e3378b3048dce648138a042012ca42eb21d59b4e7ab68abffff8c

baseline-stages/cpu-blas/identity.json

76c69c450baed44d866848639ed8f8f9da3dc4f72f0764d62a9d255a9538bbda

baseline-stages/cpu-blas/CMakeCache.txt

9a1dc6e556b6dad28dc1d49491297e5c6a00015b44df41c435644c61c734e6ae

baseline-stages/cpu-blas/python/cgpy/_core.cpython-312-x86_64-linux-gnu.so

81f5e99ede9ebf0d62e2ffcef1b4529183ce335a9d2e713d6a2d5373c7f63702

cpu-openmp campaign and stage

File

SHA256

inputs/calibration-cpu-openmp.json

91e6dcdcc17ce724bd8289952607eab9f56fc50d727782b7e92bc81dc71bb00c

calibration-cpu-openmp/plan.json

d82172afabf4cdd391dc11789e75eccac01af4bcd99862e2f6bd14019683e797

calibration-cpu-openmp/manifest.json

0e226a0c3d9924e46545fcdc4349069b2e284663c10f1b21492e7238e141cf95

calibration-cpu-openmp/reports/cpu-openmp-cpu-dense-f64-v1-jacobi-prepared-fused1-host-standard.json

00890d2d41340b7f88db91c3eb7587c372236186d284b26b980cc1bd5949d42c

baseline-stages/cpu-openmp/source-files.json

de44d54de1c7133d20fbfb3099a8e412d489a4f5e9f59fb1e53e1d8fee2d2cf8

baseline-stages/cpu-openmp/libcg_core.a

166e96b104572f1c717186c838d08e208d3391055914d8cd24334535368ceb65

baseline-stages/cpu-openmp/compile_commands.json

2a95ea3794ac02fb5fb173180aaeeb90b9b3141ae14a89ca76a95119550743ad

baseline-stages/cpu-openmp/identity.json

6d611ecf0a45a4e5f989e57c5b34528487d6bad5ec055ae60892cb9a0f57b2bb

baseline-stages/cpu-openmp/CMakeCache.txt

24c3bf3d3d0061c9a2c5a86e5271a9dc236e524bc24010ca11046498ec139250

baseline-stages/cpu-openmp/python/cgpy/_core.cpython-312-x86_64-linux-gnu.so

17d44c0b297d33607aef6c51084304e1955b3e5961fec29409e4a13a57cb1ea1

cuda-release campaign and stage

File

SHA256

inputs/calibration-cuda-release.json

afbe3b042fdb4dfde3f433f651f970f543dd8cad75a6cf5a915d6f955eb95b7e

calibration-cuda-release/plan.json

befb348aa4161d88a0a2f5101dd350d75f6daecf5b79b659ac751256377d920a

calibration-cuda-release/manifest.json

f22ee7194333605172a49abe73d50bfea245d38ba3909a099c370b662ad9516f

calibration-cuda-release/reports/cuda-release-cuda-dense-f64-v1-jacobi-api-fused1-host-standard.json

81b5248bcf731d5e80ba27ed634e63a34a0b5f7ccc1f657eda64b15b1b4236d8

calibration-cuda-release/reports/cuda-release-cuda-dense-f64-v1-jacobi-api-fused1-device-standard.json

03296e4b86f2c31a9c0f24bf6aad8feaa986570c378add8ec43e7b5a0e1ec283

calibration-cuda-release/reports/cuda-release-cuda-dense-f64-v1-jacobi-session-fused1-host-standard.json

18604528de49057c8457f946ca8de6fc0110d5e1a5732d4cd6280ffc02e20427

calibration-cuda-release/reports/cuda-release-cuda-dense-f64-v1-jacobi-prepared-fused1-device-standard.json

49da031110a71447bc7403cb73485e53d627bef392e88873a0db5b7136ab2e7a

baseline-stages/cuda-release/source-files.json

de44d54de1c7133d20fbfb3099a8e412d489a4f5e9f59fb1e53e1d8fee2d2cf8

baseline-stages/cuda-release/libcg_core.a

b6776cf9cd999ea45206ca9b5552a747575162b87f403d5297f4e5b04a35a0a9

baseline-stages/cuda-release/compile_commands.json

416932af06bed2811f8b00140891a95f8cb170ff53428d580cd1e89be47399a6

baseline-stages/cuda-release/identity.json

3a4bdbc2df5b44e8cbdc94f7b2d4de8558d9514659a0e70dd872109e3d3fd551

baseline-stages/cuda-release/CMakeCache.txt

0a4eccd32d33c0b3f5188e7434b78bdde5d532401c33cbd64c4f6dee526cb848

baseline-stages/cuda-release/python/cgpy/_cuda.cpython-312-x86_64-linux-gnu.so

2b95332f5e7cfa622e5de9eeb109ed8d6b2798f7e9eb9f8ee7d686a9c1f71fbe

baseline-stages/cuda-release/python/cgpy/_core.cpython-312-x86_64-linux-gnu.so

84d7166a496b4190f373fbdf2ffd057af8c7ada67f53670aed9d30ba78ba764a

Fixture identities

Case

Input SHA256

dense-64

b2e274e240704161b83bff5de50cf7b8ba9e3848b098f027c1ba8b5fccf82a29

dense-256

6a6622788ec7f29156523e562995e09db8ec35032185f8c76ae6fccf2165a4a0

dense-1024

204aa61929e5e8d34102ac2d0bda5714b6f0d17f8bf4796a7f71b535a21053bb

dense-2048

8bda02fed9ccafa8c12f857ad51962b02d8c860ee96a006078aed872c40cae6c