HIP: record direct atomic PSF timings

tests/test_hip_psf.hip

@@ -46,6 +46,16 @@ int main() {
     std::free(cpu_hdr); std::free(gpu_hdr); psf_kernel_cache_destroy(&cache);
     return 1;
   }
+  HipPsfTiming timing = {};
+  if (hip_psf_sink_get_timing(sink, &timing) || timing.event_count != 3 ||
+      timing.batch_count != 1 || timing.timed_batch_count != 1 ||
+      timing.upload_seconds < 0.0 || timing.kernel_seconds < 0.0 ||
+      timing.download_seconds < 0.0) {
+    std::fputs("HIP PSF timing accounting failed\n", stderr);
+    hip_psf_sink_destroy(sink);
+    std::free(cpu_hdr); std::free(gpu_hdr); psf_kernel_cache_destroy(&cache);
+    return 1;
+  }
   double max_abs = 0.0, max_rel = 0.0;
   for (size_t i = 0; i < values; ++i) {
     const double absolute = std::fabs(cpu_hdr[i] - gpu_hdr[i]);