[pull] master from arrayfire:master

src/backend/common/DefaultMemoryManager.cpp

@@ -160,7 +160,8 @@ void *DefaultMemoryManager::alloc(bool user_lock, const unsigned ndims,
         if (!this->debug_mode) {
             // FIXME: Add better checks for garbage collection
             // Perhaps look at total memory available as a metric
-            if (getMemoryPressure() > getMemoryPressureThreshold()) {
+            if (current.lock_bytes >= current.max_bytes ||
+                current.total_buffers >= this->max_buffers) {
                 this->signalMemoryCleanup();
             }
 

src/backend/cpu/queue.hpp

@@ -69,7 +69,8 @@ class queue {
 #ifndef NDEBUG
         sync();
 #else
-        if (getMemoryPressure() > getMemoryPressureThreshold() || count >= 25) {
+        if (getMemoryPressure() >= getMemoryPressureThreshold() ||
+            count >= 25) {
             sync();
         }
 #endif

src/backend/cuda/Array.cpp

@@ -248,7 +248,7 @@ kJITHeuristics passesJitHeuristics(Node *root_node) {
     // A lightweight check based on the height of the node. This is an
     // inexpensive operation and does not traverse the JIT tree.
     if (root_node->getHeight() > 6 ||
-        getMemoryPressure() > getMemoryPressureThreshold()) {
+        getMemoryPressure() >= getMemoryPressureThreshold()) {
         // The size of the parameters without any extra arguments from the
         // JIT tree. This includes one output Param object and 4 integers.
         constexpr size_t base_param_size =

src/backend/opencl/Array.cpp

@@ -296,7 +296,7 @@ kJITHeuristics passesJitHeuristics(Node *root_node) {
         return kJITHeuristics::TreeHeight;
     }
 
-    bool isBufferLimit = getMemoryPressure() > getMemoryPressureThreshold();
+    bool isBufferLimit = getMemoryPressure() >= getMemoryPressureThreshold();
     auto platform      = getActivePlatform();
 
     // The Apple platform can have the nvidia card or the AMD card