SWDEV-301667 - Improve logging

Change-Id: Ifa6da876b85cb503967cf09aac6d477b10db8e63


[ROCm/clr commit: feca11d5e3]

projects/clr/rocclr/device/rocm/rocblit.cpp

@@ -1812,12 +1812,15 @@ bool KernelBlitManager::writeBuffer(const void* srcHost, device::Memory& dstMemo
   bool result = false;
 
   if (dev().info().largeBar_ && size[0] <= kMaxH2dMemcpySize) {
-    if ((dstMemory.owner()->getHostMem() == nullptr) && (dstMemory.owner()->getSvmPtr() != nullptr)) {
+    if ((dstMemory.owner()->getHostMem() == nullptr) &&
+        (dstMemory.owner()->getSvmPtr() != nullptr)) {
       // CPU read ahead, hence release GPU memory
+      ClPrint(amd::LOG_DEBUG, amd::LOG_COPY, "Host memcpy for map wait_event");
       gpu().releaseGpuMemoryFence();
       char* dst = reinterpret_cast<char*>(dstMemory.owner()->getSvmPtr());
       std::memcpy(dst + origin[0], srcHost, size[0]);
-      // Set hasPendingDispatch_ flag. Then releaseGpuMemoryFence() will use barrier to invalidate cache
+      // Set hasPendingDispatch_ flag. Then releaseGpuMemoryFence() will use barrier
+      // to invalidate cache
       gpu().hasPendingDispatch();
       gpu().releaseGpuMemoryFence();
       return true;

projects/clr/rocclr/device/rocm/rocvirtual.cpp

@@ -146,7 +146,7 @@ void Timestamp::checkGpuTime() {
         start = std::min(time.start, start);
         end = std::max(time.end, end);
         ClPrint(amd::LOG_INFO, amd::LOG_SIG, "Signal = (0x%lx), start = %ld, "
-          "end = %ld", it->signal_.handle, start, end);
+          "end = %ld time taken= %ld ns", it->signal_.handle, start, end, end - start);
       }
       it->done_ = true;
     }