The algorithm of anisotropic scattering is reproduced from the code mentioned in [Kontar et. al. ApJ 2019]
The major computational part is done with the help of PyTorch, which is known as an excellent machine learning framework, but also a powerful tool for numerical computation, the computaion can be done with either CPU or GPU. Also the best feature is the autodiff, which carried out most of the differential works in the computaion.
Note: this is done with:
PyTorch==1.5.0
python>=3.8.3
mpi4py (for parallelize)
| Variable Name | Info |
|---|---|
dev_u |
Device decalimer, 'cuda' for GPU calculation,'cpu' for cpu |
step_N |
The number of step, complete simulation set as '-1' to let the program decide how many steps to run |
collect_N |
The steps to be collected (for debug and dev) |
photon_N |
The number of simulated photons, (recommend: GPU: 500000, CPU: 10000) |
when photon_N is smaller than 10000, CPU is faster than GPU due to the memory exchange
| Variable Name | var.shape |
|---|---|
r_vec |
[3,photon_N] |
k_vec |
[3,photon_N] |
r_vec_collect |
[t_N,3,photon_N] |
rr_cur |
[photon_N] |
tau |
[photon_N] |
For multiple GPU, try bash parallel or python multiprocessing to parallel:
cat params.input | parallel -j4 'CUDA_VISIBLE_DEVICES=$(({%} - 1)) python {} &> {#}.out'- Level 0 (lv0) x,y,z kx,ky,kz, t at sampled time points. (huge in size but contains every thing, all information of all checkpoints, just for test cases, too large)
| Variable Name |
|---|
| steps_N |
| collect_N |
| photon_N |
| start_r |
| start_theta |
| start_phi |
| f_ratio |
| epsilon |
| anis |
| asym |
| omega0 |
| freq0 |
| t_collect |
| tau |
| r_vec_collect_local |
| k_vec_collect_local |
| tau_collect_local |
- Level 1 (lv1) x,y,z kx,ky,kz at arrival shell (single time point, recommended form)
| Variable Name |
|---|
| steps_N |
| collect_N |
| photon_N |
| start_r |
| start_theta |
| start_phi |
| f_ratio |
| epsilon |
| anis |
| asym |
| omega0 |
| freq0 |
| r_vec_stat_avail |
| k_vec_stat_avail |
| t_reach_stat_avail |
| tau_stat_avail |
| r_vec_0 |
| k_vec_0 |
| dk_refr_avail |
| dk_scat_avail |
| dkx_refr_avail |
| dky_refr_avail |
| dkz_refr_avail |
| dkx_scat_avail |
| dky_scat_avail |
| dkz_scat_avail |
- Level 2 (lv2) im_x, im_y, tretraced and observed x,y position estimated (TBD)
- Level 3 (lv3) duration and source size (TBD)
pprofile --exclude-syspath sunRay_v0 > a_gpu.txtGPU tends to be faster when the number of photon photon_N >100k.