added explanation notes to syncedmem

include/caffe/syncedmem.hpp

@@ -9,32 +9,19 @@
 
 namespace caffe {
 
-
-#if 0
-
-// This chunk of code should be used when one has a machine that does not have
-// GPU, thus cannot be used if we just want to distribute a single binary.
-
-inline void CaffeMallocHost(void** ptr, size_t size) {
-  if (Caffe::mode() == Caffe::GPU) {
-    CUDA_CHECK(cudaMallocHost(ptr, size));
-  } else {
-    *ptr = malloc(size);
-  }
-}
-
-inline void CaffeFreeHost(void* ptr) {
-  if (Caffe::mode() == Caffe::GPU) {
-    CUDA_CHECK(cudaFreeHost(ptr));
-  } else {
-    free(ptr);
-  }
-}
-
-#else
-
-// This chunk of code is safer, but may not be as fast as the cuda pinned memory
-// version.
+// Theoretically, CaffeMallocHost and CaffeFreeHost should simply call the
+// cudaMallocHost and cudaFree functions in order to create pinned memory.
+// However, those codes rely on the existence of a cuda GPU (I don't know
+// why that is a must since allocating memory should not be accessing the
+// GPU resorce, but it just creates an error as of Cuda 5.0) and will cause
+// problem when running on a machine without GPU. Thus, we simply define
+// these two functions for safety and possible future change if the problem
+// of calling cuda functions disappears in a future version.
+// 
+// In practice, although we are creating unpinned memory here, as long as we
+// are constantly accessing them the memory pages almost always stays in
+// the physical memory (assuming we have large enough memory installed), and
+// does not seem to create a memory bottleneck here.
 
 inline void CaffeMallocHost(void** ptr, size_t size) {
   *ptr = malloc(size);
@@ -44,7 +31,6 @@ inline void CaffeFreeHost(void* ptr) {
   free(ptr);
 }
 
-#endif  // code to define CaffeMallocHost and CaffeFreeHost
 
 class SyncedMemory {
  public: