[Starpu-devel] [BUG] tests/datawizard/reclaim.c: hangs forever in task_wait_for_all when using OpenCL.

[Cyril Roelandt <cyril.roelandt@inria.fr>]

Hello,

I am currently trying to write OpenCL versions of our CUDA codelets for 
all the programs found in examples/ and tests/.

It is going well so far, but I jave just stumbled upon the weirdest bug 
in tests/datawizard/reclaim.c. I applied the following patch :

--- tests/datawizard/reclaim.c  (revision 6089)
+++ tests/datawizard/reclaim.c  (working copy)
@@ -34,7 +34,7 @@
 #  define BLOCK_SIZE (64*1024*1024)
 #endif

-static unsigned ntasks = 1000;
+static unsigned ntasks = 1;

 #ifdef STARPU_HAVE_HWLOC
 static uint64_t get_total_memory_size(void)
@@ -54,9 +54,9 @@

 static struct starpu_codelet dummy_cl =
 {
-        .where = STARPU_CPU|STARPU_CUDA,
        .cpu_funcs = {dummy_func, NULL},
        .cuda_funcs = {dummy_func, NULL},
+       .opencl_funcs = {dummy_func, NULL},
        .nbuffers = 3,
        .modes = {STARPU_RW, STARPU_R, STARPU_R}
 };

This patch is quite simple. It is worth noting that "dummy_func" is an 
empty function: it could not be any dummier. I have changed the value of 
ntasks because otherwise, the test takes quite a lot of time to run.

When running this test on a CPU, everything works as expected :

$ STARPU_NCPUS=1 STARPU_NCUDA=0 STARPU_NOPENCL=0 time -p 
./tests/datawizard/reclaim
...
real 2.46
user 0.02
sys 1.98

With CUDA, things start to take an awful lot of time :

$ STARPU_NCPUS=0 STARPU_NCUDA=1 STARPU_NOPENCL=0 time -p 
./tests/datawizard/reclaim
...
real 46.55
user 33.84
sys 15.13

With OpenCL, the program just hangs for ever: the program is stuck in 
starpu_task_wait_for_all().

I've run all these commands on Hannibal.

This reminds me of the bug Ludovic is currently struggling with : 
starpu_task_wait_for_all() never returns, and he uses OpenCL kernels. 
Ludovic, iirc, you started to debug this issue, so, could you tell us 
more about it ?

Cyril.