AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Changing single sync mode to time all iterations instead of just last

Browse Source
This commit is contained in:
Gilbert Lee
2019-12-20 17:08:39 -08:00
parent 000bce6f27
commit e5074ce94d
1 changed files with 8 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/TransferBench/TransferBench.cpp
+8 -5
View File
@@ -28,7 +28,9 @@ THE SOFTWARE.
#include <cstdlib> #include <cstdlib>
#include <set> #include <set>
#include <unistd.h> #include <unistd.h>
#include <map>
#include <iostream>
#include <sstream>
#include <hip/hip_runtime.h> #include <hip/hip_runtime.h>
#include "copy_kernel.h" #include "copy_kernel.h"
#include "TransferBench.hpp" #include "TransferBench.hpp"
@@ -272,7 +274,8 @@ int main(int argc, char **argv)
{ {
HIP_CALL(hipSetDevice(links[i].srcGpu)); HIP_CALL(hipSetDevice(links[i].srcGpu));
HIP_CALL(hipEventRecord(startEvents[i], streams[i])); if (!useSingleSync || iteration == 0)
HIP_CALL(hipEventRecord(startEvents[i], streams[i]));
if (useHipCall) if (useHipCall)
{ {
@@ -308,7 +311,8 @@ int main(int argc, char **argv)
gpuBlockParams[i]); gpuBlockParams[i]);
} }
} }
HIP_CALL(hipEventRecord(stopEvents[i], streams[i])); if (!useSingleSync || iteration == numIterations - 1)
HIP_CALL(hipEventRecord(stopEvents[i], streams[i]));
} }
// Synchronize per iteration, unless in single sync mode, in which case // Synchronize per iteration, unless in single sync mode, in which case
@@ -378,8 +382,7 @@ int main(int argc, char **argv)
} }
else else
{ {
if (!useSingleSync) totalGpuTime[i] /= (1.0 * numIterations);
totalGpuTime[i] /= (1.0 * numIterations);
printf("%8.3f", (linkCount[i] * numBytesPerLink / 1.0E9) / totalGpuTime[i]); printf("%8.3f", (linkCount[i] * numBytesPerLink / 1.0E9) / totalGpuTime[i]);
} }
} }