Split SAFE() macro into API-specific typed variants

Avoids incorrectly interpreting CUDA Driver Library error codes as
though they were CUDA Runtime Library error codes (the numbering is
similar, but not identical).

Also cleans up how we initialize and terminate a context for
capability checking in `mon_cross_ctx_copies`.

1 files changed, 4 insertions, 2 deletions

diff --git a/task_host_utilities.cu b/task_host_utilities.cu
index 9bb95aa..1f76080 100644
--- a/task_host_utilities.cu
+++ b/task_host_utilities.cu
@@ -4,6 +4,7 @@
 // been copied and heavily modified for use in the gpu-microbench project.
 #include <cuda_runtime.h>
 #include <errno.h>
+#include <sched.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <string.h>
@@ -191,8 +192,8 @@ static void InternalReadGPUNanoseconds(int cuda_device, double *cpu_time,
   if (!CheckCUDAError(cudaMemcpy(gpu_time, device_time, sizeof(device_time),
     cudaMemcpyDeviceToHost))) goto out;
   max_error = (cpu_end - cpu_start) / 2.0;
-  fprintf(stderr, "Time synchronized to a maximum error of +/- %f us.\n",
-    max_error * (1000.0 * 1000.0));
+  fprintf(stderr, "Time synchronized to a maximum error of +/- %f us on CPU%d.\n",
+    max_error * (1000.0 * 1000.0), sched_getcpu());
 out:
   cudaFree(device_time);
   cudaFree((void*)start_barrier);
@@ -200,6 +201,7 @@ out:
   cudaFree((void*)ready_barrier);
 }
 
+// Returns 0 on success, 1 on error
 int GetHostDeviceTimeOffset(int cuda_device, double *host_seconds,
   uint64_t *gpu_nanoseconds) {
   uint64_t *shared_gpu_time = NULL;