CodeUniverse · package
SHTns
all-spectral-coefficient gradient of a scalar spatial objective
AD speed-up on the same differentiation workload (against finite differences) Speed-up: 54.6×
High-dimensional demo
all-spectral-coefficient gradient of a scalar spatial objective
Supplied demo scenario wording.
FD components requested 1 – 256 · AD components 1024
View demo measurements
| Measure | Value |
|---|---|
| AD median (s) | 0.00013879092875868082 |
| FD median (s) | 0.00757716700900346 |
| FD / AD | 54.594108395787636 |
| Max scaled error | 3.432158383054329e-7 |
| AD components | 1024 |
| FD components requested | 256 |
| AD parameters | 256 |
| Repeats | 7 |
| FD components requested | AD median (s) | FD median (s) | FD / AD | Max scaled error | AD components |
|---|---|---|---|---|---|
| 1 | 1024 | 0.00013620802201330662 | 0.00003162503708153963 | 0.2321818980562692 | 2.0945844608301974e-10 |
| 2 | 1024 | 0.00013695808593183756 | 0.000060999998822808266 | 0.445391729942599 | 2.756671389349418e-9 |
| 4 | 1024 | 0.00013741699513047934 | 0.00012058403808623552 | 0.8775045471758374 | 1.9546396225773617e-8 |
| 8 | 1024 | 0.00013554200995713472 | 0.00023849995341151953 | 1.7596017167440954 | 8.227034368004986e-8 |
| 16 | 1024 | 0.0001349169760942459 | 0.00047450000420212746 | 3.5169777587563678 | 8.227034368004986e-8 |
| 32 | 1024 | 0.00013349996879696846 | 0.0009456659900024533 | 7.083642030213925 | 1.259285234422869e-7 |
| 64 | 1024 | 0.00013504107482731342 | 0.001889208098873496 | 13.989877533849313 | 1.259285234422869e-7 |
| 128 | 1024 | 0.00013870897237211466 | 0.0037465000059455633 | 27.009788493673106 | 3.432158383054329e-7 |
| 256 | 1024 | 0.00013879092875868082 | 0.00757716700900346 | 54.594108395787636 | 3.432158383054329e-7 |
One-direction performance audit
View audit measurements
| Demo scenario | Problem size | Repeats | AD median (s) | FD median (s) | FD / AD | Relative error | Supplied verdict |
|---|---|---|---|---|---|---|---|
| scalar_transform_direction | 16x12 | — | 0.000012082979083 | 0.0000073750270531 | 0.610364961 | 1.0715e-10 | — |
FD/AD > 1 means automatic differentiation is faster; FD/AD < 1 means the finite-difference comparison is faster. One direction only, measured inside the package's own case.