For same time step size GPU explicit solver diverges but CPU version does not

[luchete80]

[luchete80]

Same thing with finer mesh

[luchete80]

For 1320 particles

Max disp 2.66227e-05, 2.66227e-05, 0
Step 81, Min step size 7.27353e-07
--------------------------- END STEP, Time8.2e-05, --------------------------
Step 82, Min step size 9.34472e-07
--------------------------- END STEP, Time8.3e-05, --------------------------
Step 83, Min step size 9.69382e-07
--------------------------- END STEP, Time8.4e-05, --------------------------
Step 84, Min step size 9.3563e-07
--------------------------- END STEP, Time8.5e-05, --------------------------
Step 85, Min step size 8.96999e-07
--------------------------- END STEP, Time8.6e-05, --------------------------
Step 86, Min step size 9.88032e-07
--------------------------- END STEP, Time8.7e-05, --------------------------
Step 87, Min step size 9.53742e-07
--------------------------- END STEP, Time8.8e-05, --------------------------
Step 88, Min step size 4.8974e-07
--------------------------- END STEP, Time8.9e-05, --------------------------
Step 89, Min step size 9.74918e-07
--------------------------- END STEP, Time9e-05, --------------------------
Step 90, Min step size 9.15505e-07
--------------------------- END STEP, Time9.1e-05, --------------------------
Time 9.1e-05, GPU time 0.239

[luchete80]

https://stackoverflow.com/questions/14406364/different-results-for-cuda-addition-on-host-and-on-gpu

https://forums.developer.nvidia.com/t/cpu-and-gpu-floating-point-anomaly/31100

https://developer.download.nvidia.com/assets/cuda/files/NVIDIA-CUDA-Floating-Point.pdf

The nvcc compiler switch, --fmad (short name: -fmad), to control the contraction of
floating-point multiplies and add/subtracts into floating-point multiply-add
operations (FMAD, FFMA, or DFMA) has been added:
--fmad=true and --fmad=false enables and disables the contraction respectively.
This switch is supported only when the --gpu-architecture option is set with
compute_20, sm_20, or higher. For other architecture classes, the contraction is
always enabled.
The --use_fast_math option implies --fmad=true, and enables the contraction.

[luchete80]

New KickDrift Algorithm:
GPU step size 0.2:

---

Time 0.0101001, GPU time 221.382
Current time step: 7.18303e-07
Forces calc: 0
Stresses calc: 2.733
Max disp: 0.0147366, 0.0147429, 0.00209899
Max pl_strain: 0.262628
Total steps: 14061, time spent 221.388000
Program ended.

[luchete80]

Diverging with Step Size with CFL 0.4 (in CPU converges)

[luchete80]

[luchete80]

In new compilation, it works:

Max pl_strain: 0.0844634

Time 0.00500011, GPU time 105.279
Current time step: 7.18303e-07
Forces calc: 0
Stresses calc: 1.36
Max disp: 0.0048783, 0.00488042, 0.000351309
Max pl_strain: 0.0880185

[luchete80]

Maybe is with something related to flag mentioned early, now is working on CFL 0.4:

---

Time 0.0100002, GPU time 218.864
Current time step: 7.18303e-07
Forces calc: 0
Stresses calc: 2.897
Max disp: 0.0145451, 0.0145513, 0.00206436
Max pl_strain: 0.259194

Time 0.0101001, GPU time 221.082
Current time step: 7.18303e-07
Forces calc: 0
Stresses calc: 2.929
Max disp: 0.0147366, 0.0147429, 0.00209899
Max pl_strain: 0.262628
Total steps: 14061, time spent 221.086000
Program ended.

[luchete80]

ERROR