allocate CPUs to take care of the gpu streams and do not use them as CPU workers + use a single gpu for the gpu_partitionong example

examples/sched_ctx/gpu_partition.c

@@ -109,8 +109,8 @@ int main(int argc, char **argv)
 	int ncuda = 0;
 	int gpu_devid = -1;
 
-#warning temporary fix: skip test as cuda computation fails
-	return 77;
+/* #warning temporary fix: skip test as cuda computation fails */
+/* 	return 77; */
 
 #ifndef STARPU_HAVE_SETENV
 	return 77;
@@ -118,6 +118,7 @@ int main(int argc, char **argv)
 	/* Have separate threads for streams */
 	setenv("STARPU_CUDA_THREAD_PER_WORKER", "1", 1);
 	setenv("STARPU_NWORKER_PER_CUDA", "2", 1);
+	setenv("STARPU_NCUDA", "1", 1);
 #endif
 
 	/* Initialize StarPU */
@@ -175,7 +176,7 @@ int main(int argc, char **argv)
 	int ncpus = starpu_cpu_worker_get_count();
 	int workers[ncpus+nstreams];
 	starpu_worker_get_ids_by_type(STARPU_CPU_WORKER, workers, ncpus);
-
+	
 	int sched_ctxs[nstreams];
 	int nsms[nstreams];
 	nsms[0] = 6;

src/core/topology.c

@@ -1443,8 +1443,13 @@ _starpu_init_machine_config(struct _starpu_machine_config *config, int no_mp_con
             mpi_ms_busy_cpus = 1; /* we launch one thread to control all slaves */
 #endif
 #endif /* STARPU_USE_MPI_MASTER_SLAVE */
-
-			unsigned already_busy_cpus = mpi_ms_busy_cpus + mic_busy_cpus + topology->ncudagpus
+	    unsigned cuda_busy_cpus = 0;
+#if defined(STARPU_USE_CUDA)
+	    cuda_busy_cpus = th_per_stream ? (nworker_per_cuda * topology->ncudagpus) : 
+		    topology->ncudagpus;
+#endif
+			unsigned already_busy_cpus = mpi_ms_busy_cpus + mic_busy_cpus 
+				+ cuda_busy_cpus
 				+ topology->nopenclgpus + topology->nsccdevices;
 
 			long avail_cpus = (long) topology->nhwcpus - (long) already_busy_cpus;