Merge topic 'fix-ManagedMemory-issue'

b530a5ce

 Fix issue with Managed Memory for 0 size arrays
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Robert Maynard <robert.maynard@kitware.com>
Merge-request: !1033

vtkm/cont/cuda/internal/CudaAllocator.cu

@@ -33,6 +33,10 @@ static bool IsInitialized = false;
 
 // True if all devices support concurrent pagable managed memory.
 static bool ManagedMemorySupported = false;
+
+// Avoid overhead of cudaMemAdvise and cudaMemPrefetchAsync for small buffers.
+// This value should be > 0 or else these functions will error out.
+static std::size_t Threshold = 1 << 20;
 }
 
 namespace vtkm
@@ -94,6 +98,12 @@ bool CudaAllocator::IsManagedPointer(const void* ptr)
 void* CudaAllocator::Allocate(std::size_t numBytes)
 {
   CudaAllocator::Initialize();
+  // When numBytes is zero cudaMallocManaged returns an error and the behavior
+  // of cudaMalloc is not documented. Just return nullptr.
+  if (numBytes == 0)
+  {
+    return nullptr;
+  }
 
   void* ptr = nullptr;
   if (ManagedMemorySupported)
@@ -115,7 +125,7 @@ void CudaAllocator::Free(void* ptr)
 
 void CudaAllocator::PrepareForControl(const void* ptr, std::size_t numBytes)
 {
-  if (IsManagedPointer(ptr))
+  if (IsManagedPointer(ptr) && numBytes >= Threshold)
   {
 #if CUDART_VERSION >= 8000
     // TODO these hints need to be benchmarked and adjusted once we start
@@ -128,7 +138,7 @@ void CudaAllocator::PrepareForControl(const void* ptr, std::size_t numBytes)
 
 void CudaAllocator::PrepareForInput(const void* ptr, std::size_t numBytes)
 {
-  if (IsManagedPointer(ptr))
+  if (IsManagedPointer(ptr) && numBytes >= Threshold)
   {
 #if CUDART_VERSION >= 8000
     int dev;
@@ -143,7 +153,7 @@ void CudaAllocator::PrepareForInput(const void* ptr, std::size_t numBytes)
 
 void CudaAllocator::PrepareForOutput(const void* ptr, std::size_t numBytes)
 {
-  if (IsManagedPointer(ptr))
+  if (IsManagedPointer(ptr) && numBytes >= Threshold)
   {
 #if CUDART_VERSION >= 8000
     int dev;
@@ -158,7 +168,7 @@ void CudaAllocator::PrepareForOutput(const void* ptr, std::size_t numBytes)
 
 void CudaAllocator::PrepareForInPlace(const void* ptr, std::size_t numBytes)
 {
-  if (IsManagedPointer(ptr))
+  if (IsManagedPointer(ptr) && numBytes >= Threshold)
   {
 #if CUDART_VERSION >= 8000
     int dev;