Rewrite clock rate synchronization code to increase accuracy

Mirrors architecture used for instantaneous clock synchronization.

The PPM (parts-per-million) error after a two-second
synchronization period (using results from new implementation after
a 100-second synchronization period as ground truth, as after 100
seconds, the new and old implementation disagree by at most 2 PPM).

GPU (machine name)              Before  After
------------------------------  ------  ------
GTX 1060 3 GiB (jbakita-old):   12      1
GTX 970 (jbakita-old):          9       2
GTX 1080 Ti (yamaha):           5       1
RTX 6000 Ada (yamaha):          4       2
(An error of 1 PPM is 1 microsecond per second.)

Also modify `preemption_logger` to print the rate of skew as PPM,
rather than just as a multiplier.

1 files changed, 7 insertions, 4 deletions

diff --git a/preemption_logger.cu b/preemption_logger.cu
index 89b348a..ec3c22e 100644
--- a/preemption_logger.cu
+++ b/preemption_logger.cu
@@ -90,10 +90,11 @@ int main(int argc, char **argv) {
         // Synchronize the GPU and CPU clocks (if requested) using the utilities
         // from task_host_utilities.cu.
         if (!skip_conversion) {
-                // Note that this appears to almost always be 1.0 on recent GPUs
-                d2h_scale = GetGPUTimerScale(0);
+                // The skew is between -13 and 60 microseconds per second on the GTX
+                // 1080, GTX 1060 3 GiB, GTX 970, RTX 6000 Ada, and GTX 1080 Ti.
+                d2h_scale = InternalGetGPUTimerScale(0);
                 InternalReadGPUNanoseconds(0, &host_s, &dev_ns);
-                if (host_s == 0 && !dev_ns) {
+                if (d2h_scale == -1 || (host_s == 0 && !dev_ns)) {
                         fprintf(stderr, "Unable to synchronize time with the GPU. Aborting...\n");
                         return 1;
                 }
@@ -153,7 +154,9 @@ int main(int argc, char **argv) {
                 fprintf(stderr, "(%d) Aprox launch overhead: %ld ns\n", pid, launch_oh);
                 fprintf(stderr, "(%d) CPU clock - GPU clock: %ld tick gap\n", pid,
                     (long)s2ns(host_s) - dev_ns);
-                fprintf(stderr, "(%d) 1 CPU tick/1 GPU tick: %.3f\n", pid, d2h_scale);
+                fprintf(stderr, "(%d) After 1 second, the GPU clock is %.f us %s (%.9fx d2h)\n",
+                    pid, fabs((d2h_scale - 1) * 1e6),
+                    d2h_scale > 1 ? "behind" : "ahead", d2h_scale);
         }
 
         return 0;