I have successfully cross-compiled ACADOS for VxWorks 6.9, a 32-bit system running on PPC CPU. Here are the settings
So
The target is set to GENERIC
EXT_DEP is turned off
LA uses REFERENCE
MF uses COLMAJ
The compilation is successful, despite a few warnings. Then I successfully compiled locally the c code of the bicycle model generated from the Python interface. For simplicity, I tested the integrator first.
When I try to run the executable, I got the following output
So basically the code is not working. Ok, I start debug, and found that the error came from the struct mem in the integrator.
When I try to give values to the mem->time_ad, mem->time_la and mem->time_sim, the error shows. Then I found that these three double variables are not created, or their memory is not assigned properly.
Finally, I simply comment these three lines, and the integrator works!
My question is, why is the case? There seems memory alignment or assignment issues, but weirdly everything works fine on Ubuntu 64bit system, and on Speedgoat.
To further locate the problem, I test the OCP example as well, and the code stuck at model->scaling=1.0
The similarity of these two tests is that: both sim_irk_memory and model struct contains a few double pointers and a few double variables. The pointers work fine, but the double variables are crashed. I guess, this is the problem of align_char_to(8, &c_ptr), or any thing related to this?
Hence, it seems that an initial memory alignment is necessary in my case.
Although this solves my issue for the integrator test, further tests on OCP example show that I need to modify dozens of functions, which makes this workaround tedious. Is there better way to to it?
Thank you for the reply. The guide you gave is very helpful.
However, after adapting the memory alignment guide to our architecture, we still got some errors. The errors indicate that HPIPM structs have memory problems, and this kind of struct is not explained in the guide you provided.
Specifically, we found that the first error lied in the function
Can you try to add an align here?
i.e. before d_ocp_qp_ipm_ws, which has some doubles at the beginning.
Also, HPIPM has been successfully tested on a 32bit ARM, but it might be that there are issues left for PPC. It might be also related to the fact that PPC is big endian is big endian by default.
I hope this helps you to fix it and that you contribute the fix!
However, we went through the code deeper and find out mem->hpipm_workspace->stat address can be printed out, and mem->hpipm_workspace->stat_m has a value of 17.
When we tried to print mem->hpipm_workspace->stat[0], the program stoped. Hence, it is obvious that the memory of this vector is not assigned properly.
We then checked this vector in x_ocp_qp_ipm.c (in function void OCP_QP_IPM_WS_CREATE), and found the codes below
// double/float stuff
REAL *d_ptr = (REAL *) sv_ptr;
workspace->stat = d_ptr;
int stat_m = 17;
d_ptr += stat_m*(1+arg->stat_max);
So, stat_m=17 is confirmed, and clearly the vector “stat” has been assigned memory here, but somehow not successfully when returned to function void *ocp_qp_hpipm_memory_assign
Finally, we tried a very direction solution, i.e. adding manually the memory for this vector in the function void *ocp_qp_hpipm_memory_assign as
Now I’m not sure if this can be a PR (does it work for other architectures?), and why the memory of a single vector is not assigned inside HPIPM create function correctly so that we have to do it manually in ACADOS create function.
I am also struggling with memory erros on a QNX with PPC machine (dSPACE) and tried this fix. However I get the following error, when applying your fix. Do we have different code versions?
‘ocp_qp_hpipm_opts {aka struct ocp_qp_hpipm_opts_}’ has no member named ‘stat_max’
assign_and_advance_double(17*(1+opts->stat_max), &ipm_workspace->stat, &c_ptr);
