Marek Kaluba

Dear @bodono I managed to get access to a gpu-enabled node and run some tests there; * a simple `make test_gpu` which results in ``` ~/local/scs$ ldd ./out/run_tests_gpu_indirect linux-vdso.so.1 (0x00007fff935d2000)...

hmm, precisely the same problem happens if I compile with ``` BLASLDFLAGS="-L${JULIA_BLAS_PATH} -lopenblas" SCSFLAGS="USE_OPENMP=0 BLAS32=1 DLONG=0" make purge CUDA_PATH="${CUDA_PATH}" make -j4 $SCSFLAGS BLASLDFLAGS="${BLASLDFLAGS}" test_gpu ``` here is a gist from...

@bodono could you have a look at this problem: https://cloud.impan.pl/s/MX5oBX0lHb5LJl2 It's the same problem that you obtain through this code: ```julia let T = SCS.GpuIndirectSolver A = [ 1.0 1.0...

this is the script I use to compile `scs` ``` script = raw""" cd $WORKSPACE/srcdir/scs* flags="DLONG=0 BLAS32=1 USE_OPENMP=0 INDIRECT=1" blasldflags="-L${libdir} -lopenblas" CUDA_PATH=$prefix/cuda make BLASLDFLAGS="${blasldflags}" ${flags} out/libscsgpuindir.${dlext} mkdir -p ${libdir} cp...

`DINDIRECT=1` results in the same log

have you tried to set `acceleration_lookback=0`?

and can you gist the scs logs from those runs?

with `acceleration_lookback=0` scs seems to be progressing, albeit slowly (just to check that it actually does: try giving a larger `max_iters`, e.g. 2_000_000 or so). `acceleration_lookback=1` is numerical instability in...

Sure, is `cbf` (from https://github.com/mlubin/ConicBenchmarkUtilities.jl) format ok? (the way I construct the problem is quite complicated)

The problem data is here: https://cloud.impan.pl/s/ArwRGZuOo0YNMqc In julia You need to execute: ```julia using MathProgBase using ConicBenchmarkUtilities using SCS solver = SCSSolver(alpha=1.5, acceleration_lookback=2) dat = readcbfdata("AutF5.cbf.gz") c, A, b, con_cones,...