./out/demo_socp_gpu fails to solve its problem #180
Comments
|
Thanks for posting. I am unable to reproduce this, when I run the command I get: It might be the case that you are missing the gpu fixes I submitted here: 13e675d. I did not cut a new release / tag with those fixes. Is that the issue? By the way, you can better test the gpu using: |
|
I'm on make purge
make test_gpuas advised and then test it with no fancy options, no julia-shipped blas ;) |
|
That's strange, I cannot reproduce this on the only gpu machine I have access to. Can you try disabling the AA? You can do it by changing Here's what my ldd looks like, I don't see any major differences to yours: Can you try running it likely won't help (and is very noisy for gpus) but just in case. |
|
I disabled AA but it changed just the numerical values in the log, not the behaviour; I think the problem is somewhere in problem generation (before scs), since the header looks like this: i.e. first non positive eps_abs and then a problem with 2 variables and 4 constraints? |
|
That's just the output of the first test which is testing data validation and is working correctly. You will see the same if you run the non gpu tests with |
|
I have got the same problem with @kalmarek . |
yeah, maybe I should try to compare with |
|
@bodono: so I set seem to go away if i replace macro expanded |
|
Ok, can you try with The macro itself has an error check when VERBOSITY>0 (see here), which is why the error goes away when you replace it (although it does suggest that only that line is broken, which is strange). |
|
Sorry, false alarm. |
|
Even with I'm not sure how to test that my CUDA/cublas is installed properly? |
|
Can you try setting |
|
I set it to 1 but I get a similar behavior (though no errors). I also checked that |
|
This is so strange, I don't understand what's happening here at all and I can't reproduce this behavior on my gpu machine. If you really want to get to the bottom of this then I'm happy to get on a call and we can debug together manually on your machine. |
|
Thanks! I asked for the access to a nvidia gpu at my institution; If I can reproduce it there I'll get back to you! |
|
Dear @bodono
runs just fine (11 out of 11 tests passed).
BLASLDFLAGS="-L${JULIA_BLAS_PATH} -lopenblas64_"
make purge
make -j4 $SCSFLAGS BLASSUFFIX="_64_" BLAS64=1 DLONG=0 BLASLDFLAGS="${BLASLDFLAGS}" test_gpuwhich results in I get a failure:
|
|
Hmmm, if the blas you're using is 64 bit it might be tricky to get everything to work with a GPU which (usually) expects 32 bit integers. |
|
hmm, precisely the same problem happens if I compile with here is a gist from build, tests and ldd. OpenBLAS is the |
|
Looks like the tests are passing except for hs21, which is probably just because the numerics are slightly different on the GPU and it's producing a bad flag. |
|
@bodono could you have a look at this problem: It's the same problem that you obtain through this code: let T = SCS.GpuIndirectSolver
A = [
1.0 1.0 0.0 0.0 0.0
0.0 1.0 0.0 0.0 1.0
0.0 0.0 1.0 1.0 1.0
-1.0 0.0 0.0 0.0 0.0
0.0 -1.0 0.0 0.0 0.0
0.0 0.0 -1.0 0.0 0.0
0.0 0.0 0.0 -1.0 0.0
0.0 0.0 0.0 0.0 -1.0
]
m, n = Int32.(size(A))
args = (
m = m,
n = n,
A = A,
P = zeros(n, n),
b = [5.0, 3.0, 9.0, 0.0, 0.0, 0.0, 0.0, 0.0],
c = -[3.0, 4.0, 4.0, 9.0, 5.0],
z = 0,
l = 8,
bu = Float64[],
bl = Float64[],
q = Int32[],
s = Int32[],
ep = 0,
ed = 0,
p = Float64[],
)
solution = SCS.scs_solve(T, args..., max_iters=200, write_data_filename="simple_problem.scs")
@test isapprox(solution.x' * args.c, -99.0; rtol = 1e-4)
endThis is easily solvable by the this is what I get here: |
|
Did you compile with the |
|
this is the script I use to compile |
|
|
|
The error message When the INDIRECT flag is set SCS does the additional computation to generate a good warm-start and a sensible tolerance for the indirect system: Line 366 in f2da64d Otherwise the tolerance is set to -1.0, which is an invalid tolerance: Line 361 in f2da64d And that trips a warning from the indirect system solvers (should probably error out): scs/linsys/gpu/indirect/private.c Line 474 in 8ca0377 When that flag is not set SCS skips that computation for speed. |
|
Hmmm, actually this is likely something to do with the GPU solver specifically. There is some issue in there that only trips on some GPUs that I have run into before. It's probably something to do with type sizes that I have not been able to figure out. I would probably recommend shelving the GPU solver for now, the MKL one is typically faster anyway. |
|
Try the following patch. I got all the tests to pass with this fix. --- a/linsys/gpu/gpu.c
+++ b/linsys/gpu/gpu.c
@@ -19,13 +19,13 @@ void SCS(accum_by_atrans_gpu)(const ScsGpuMatrix *Ag,
if (*buffer != SCS_NULL) {
cudaFree(*buffer);
}
- cudaMalloc(buffer, *buffer_size);
+ cudaMalloc(buffer, new_buffer_size);
*buffer_size = new_buffer_size;
}
CUSPARSE_GEN(SpMV)
(cusparse_handle, CUSPARSE_OPERATION_NON_TRANSPOSE, &onef, Ag->descr, x,
- &onef, y, SCS_CUDA_FLOAT, SCS_CSRMV_ALG, buffer);
+ &onef, y, SCS_CUDA_FLOAT, SCS_CSRMV_ALG, *buffer);
}
/* this is slow, use trans routine if possible */
@@ -48,13 +48,13 @@ void SCS(accum_by_a_gpu)(const ScsGpuMatrix *Ag, const cusparseDnVecDescr_t x,
if (*buffer != SCS_NULL) {
cudaFree(*buffer);
}
- cudaMalloc(buffer, *buffer_size);
+ cudaMalloc(buffer, new_buffer_size);
*buffer_size = new_buffer_size;
}
CUSPARSE_GEN(SpMV)
(cusparse_handle, CUSPARSE_OPERATION_TRANSPOSE, &onef, Ag->descr, x, &onef, y,
- SCS_CUDA_FLOAT, SCS_CSRMV_ALG, buffer);
+ SCS_CUDA_FLOAT, SCS_CSRMV_ALG, *buffer);
}
/* This assumes that P has been made full (ie not triangular) and uses the |
|
@syockit Thanks for this! I applied the patch and it worked! Do you want to turn this into a PR? The only problem I had was an erroneous 'infeasible' certificate on |
|
@bodono It's a hassle for me to set up a fork right now, so please apply the commit on your side. You're right, I got the same infeasible certificate on the tests you mentioned. I missed that yesterday. And tightening |
|
Sure, no problem @syockit , thanks for sending in the patch! |
I presume this issue can be closed after #251 is merged |
Specifications
masterat 5be0e16Description
scs fails at solving
./out/demo_socp_gpu 1000 0.5 0.5 1How to reproduce
linking against julia openblas:
then running it via
Additional information
similarly compiled direct and indirect solvers (cpu) work just fine
Output
The text was updated successfully, but these errors were encountered: