Added documentation of memory types (#770)

docs_input/basics/memory.rst

@@ -0,0 +1,19 @@
+.. _memory:
+
+Memory
+======
+
+MatX allows tensors to be allocated in several different spaces either corresponding to physical or logical
+allocations. The space where the memory is allocated dictates where a user is allowed to access the memory.
+For example, allocating device memory on some system is not accessible by the CPU and will result in a SEGFAULT
+when trying to access.
+
+The type of allocation may behave differently across systems. For example, on a Grace-Hopper system (GH200) 
+standard host memory from malloc is accessible from the GPU, but may not be on other platforms. Some types
+may not be available in certain environments. On WSL2 CUDA Unified Memory (UM) or managed memory is not fully
+supported and may result in slow code or other issues.
+
+The memory type is typically chosen when creating a tensor with `make_tensor`. The memory *may* be allocated
+immediately, but it is not guaranteed. The memory is guaranteed to be available before it used used, however.
+
+.. doxygenenum:: matxMemorySpace_t

include/matx/core/allocator.h

@@ -55,12 +55,12 @@ namespace matx {
  * 
  */
 enum matxMemorySpace_t {
-  MATX_MANAGED_MEMORY,
-  MATX_HOST_MEMORY,
-  MATX_HOST_MALLOC_MEMORY,
-  MATX_DEVICE_MEMORY,
-  MATX_ASYNC_DEVICE_MEMORY,
-  MATX_INVALID_MEMORY
+  MATX_MANAGED_MEMORY,      ///< CUDA managed memory or CUDA Unified Memory (UM) from cudaMallocManaged
+  MATX_HOST_MEMORY,         ///< CUDA host-pinned memory from cudaHostAlloc
+  MATX_HOST_MALLOC_MEMORY,  ///< Host-alloced memory (pageable) from malloc
+  MATX_DEVICE_MEMORY,       ///< CUDA device memory from cudaMalloc
+  MATX_ASYNC_DEVICE_MEMORY, ///< CUDA asynchronous device memory corresponding to a stream from cudaMallocAsync
+  MATX_INVALID_MEMORY       ///< Sentinel value
 };
 
 namespace detail {