Benchmark#

The SFA of hhgsfa is checked against StrongFieldApproximation.jl by S. Carlström, a Julia code for the same model. Both codes run on exactly the same fields, in argon at 1030 nm.

Summary#

  • With every return kept, the two codes give the same harmonics, to 2 × 10⁻⁵ at 128 points per cycle and 10⁻⁷ at 512.

  • With the first return only, they differ by about 13 %, because they cut the long excursions differently. hhgsfa uses a smooth cut, which also converges faster.

  • hhgsfa is 10 to 20 times faster with numpy alone, and about 100 times faster with numba on a single core.

Differences#

Both use the Lewenstein model, the saddle-point momentum and a hydrogen-like atom without depletion. They differ in the numerics.

hhgsfa

StrongFieldApproximation.jl

Vector potential and its integrals

from the sampled field, 4th order

analytic

Sum over excursion times

trapezoid

rectangle

First return

smooth cut at one cycle

hard cut at one cycle

Parallel

yes, with numba

no

Figures#

Spectra of a 30 fs pulse at 1.5 × 10¹⁴ W/cm². The two codes lie on top of each other.

Spectra of both codes for a 30 fs pulse

Harmonics 17, 21 and 31 against the peak intensity, amplitude and phase.

Amplitude and phase of H17, H21 and H31 against the peak intensity, both codes

Accuracy against the number of points per cycle, for a 10 fs pulse.

Convergence of both codes and their difference against the number of points per cycle

Run time of the dipole.

Run time of the dipole against the number of time points, for both codes and each backend

Running the benchmark#

The scripts are in benchmarks/sfa_jl. Install the Julia reference once, then run from the root of the repository.

julia --project=env -e 'using Pkg;
    Pkg.add(url="https://github.com/jagot/StrongFieldApproximation.jl", rev="835080b");
    Pkg.add(["ElectricFields", "Unitful", "UnitfulAtomic", "NPZ"])'
JULIA_PROJECT=env uv run --extra numba --with matplotlib python benchmarks/sfa_jl/run.py

Tables#

Julia compile time 3.3 s (excluded below).

Accuracy#

window

points per cycle

hhgsfa, own convergence

julia, own convergence

between codes

between codes, phase [rad]

first return

128

5.0e-05

2.7e-02

1.2e-01

3.6e-02

first return

256

8.2e-06

1.1e-02

1.3e-01

4.0e-02

first return

512

7.5e-07

4.0e-03

1.4e-01

4.2e-02

first return

1024

reference

reference

1.4e-01

4.3e-02

every return

128

5.3e-04

5.3e-04

1.9e-05

2.7e-05

every return

256

1.6e-04

1.6e-04

1.2e-06

1.7e-06

every return

512

reference

reference

7.4e-08

1.1e-07

Run time#

Measured on 13th Gen Intel(R) Core(TM) i7-13700, 24 logical cores.

case

time points

julia, 1 thread [s]

hhgsfa numpy [s]

hhgsfa numba, 1 thread [s]

hhgsfa numba, 24 threads [s]

10 fs, first return, 128 pts/cycle

1491

0.21

0.012 (×18)

0.003 (×82)

0.001 (×292)

10 fs, first return, 256 pts/cycle

2981

0.88

0.052 (×17)

0.012 (×75)

0.003 (×316)

10 fs, first return, 512 pts/cycle

5961

3.66

0.189 (×19)

0.045 (×80)

0.011 (×330)

10 fs, first return, 1024 pts/cycle

11922

15.03

0.641 (×23)

0.160 (×94)

0.038 (×393)

10 fs, every return, 128 pts/cycle

1491

1.29

0.122 (×11)

0.019 (×67)

0.005 (×244)

10 fs, every return, 256 pts/cycle

2981

5.41

0.445 (×12)

0.072 (×75)

0.024 (×224)

10 fs, every return, 512 pts/cycle

5961

22.24

1.742 (×13)

0.221 (×101)

0.086 (×258)

Harmonic contrast, 30 fs at 1.5e14 W/cm²#

Amplitude of each harmonic over the larger of its two even neighbors.

window

code

H17

H19

H21

H31

H41

first return

hhgsfa

0.6

0.9

1.0

1.2

1.1

first return

julia

0.7

1.0

1.0

1.2

1.1

every return

hhgsfa

0.7

0.6

0.3

1.0

1.1

every return

julia

0.7

0.6

0.3

1.0

1.1