add garbage collection for AsyncOps

ROCm · Apr 18, 2016 · def3939 · def3939
1 parent e397525
commit def3939
Showing 1 changed file with 11 additions and 0 deletions.
diff --git a/lib/hsa/mcwamp_hsa.cpp b/lib/hsa/mcwamp_hsa.cpp
@@ -23,6 +23,7 @@
 #include <thread>
 #include <utility>
 #include <vector>
+#include <algorithm>
 
 #include <hsa.h>
 #include <hsa_ext_finalize.h>
@@ -66,6 +67,10 @@
 // default set as 0 (NOT print out kernel dispatch time)
 #define KALMAR_DISPATCH_TIME_PRINTOUT (0)
 
+// threshold to clean up finished kernel in HSAQueue.asyncOps
+// default set as 1024
+#define ASYNCOPS_VECTOR_GC_SIZE (1024)
+
 
 
 static const char* getHSAErrorString(hsa_status_t s) {
@@ -926,6 +931,12 @@ class HSAQueue final : public KalmarQueue
                 asyncOps[i] = nullptr;
             }
         }
+
+        // GC for finished kernels
+        if (asyncOps.size() > ASYNCOPS_VECTOR_GC_SIZE) {
+          asyncOps.erase(std::remove(asyncOps.begin(), asyncOps.end(), nullptr),
+                         asyncOps.end());
+        }
     }
 };