TA Tensor memory trace

CMakeLists.txt

@@ -144,7 +144,10 @@ if((ENABLE_GPERFTOOLS OR ENABLE_TCMALLOC_MINIMAL) AND CMAKE_SYSTEM_NAME MATCHES
   add_feature_info(Libunwind ENABLE_LIBUNWIND "Libunwind provides stack unwinding")
 endif()
 
-option(TA_TENSOR_MEM_PROFILE "Turn on instrumented profiling of TA::Tensor memory use" OFF)
+option(TA_TENSOR_MEM_TRACE "Turn on instrumented tracing of TA::Tensor memory use" OFF)
+add_feature_info(TENSOR_MEM_TRACE TA_TENSOR_MEM_TRACE "instrumented tracing of TA::Tensor memory use")
+
+option(TA_TENSOR_MEM_PROFILE "Turn on instrumented profiling of TA::Tensor memory use" ${TA_TENSOR_MEM_TRACE})
 add_feature_info(TENSOR_MEM_PROFILE TA_TENSOR_MEM_PROFILE "instrumented profiling of TA::Tensor memory use")
 
 option(TA_EXPERT "TiledArray Expert mode: disables automatically downloading or building dependencies" OFF)

INSTALL.md

@@ -417,6 +417,7 @@ support may be added.
 * `TA_SIGNED_1INDEX_TYPE` -- Set to `OFF` to use unsigned 1-index coordinate type (default for TiledArray 1.0.0-alpha.2 and older). The default is `ON`, which enables the use of negative indices in coordinates.
 * `TA_MAX_SOO_RANK_METADATA` -- Specifies the maximum rank for which to use Small Object Optimization (hence, avoid the use of the heap) for metadata. The default is `8`.
 * `TA_TENSOR_MEM_PROFILE` -- Set to `ON` to profile host memory allocations used by TA::Tensor. This causes the use of Umpire for host memory allocation. This also enables additional tracing facilities provided by Umpire; these can be controlled via [environment variable `UMPIRE_LOG_LEVEL`](https://umpire.readthedocs.io/en/develop/sphinx/features/logging_and_replay.html), but note that the default is to log Umpire info into a file rather than stdout.
+* `TA_TENSOR_MEM_TRACE` -- Set to `ON` to *trace* host memory allocations used by TA::Tensor. This turns on support for tracking memory used by `Tensor` objects; such tracking must be enabled programmatically. This can greatly increase memory consumption by the application and is only intended for expert developers troubleshooting memory use by TiledArray.
 * `TA_UT_CTEST_TIMEOUT` -- The value (in seconds) of the timeout to use for running the TA unit tests via CTest when building the `check`/`check-tiledarray` targets. The default timeout is 1500s.
 
 # Build TiledArray

src/CMakeLists.txt

@@ -189,6 +189,8 @@ TiledArray/util/bug.h
 TiledArray/util/function.h
 TiledArray/util/initializer_list.h
 TiledArray/util/logger.h
+TiledArray/util/ptr_registry.cpp
+TiledArray/util/ptr_registry.h
 TiledArray/util/random.h
 TiledArray/util/singleton.h
 TiledArray/util/threads.h

src/TiledArray/config.h.in

@@ -81,6 +81,9 @@
 /* Is TA::Tensor memory profiling enabled? */
 #cmakedefine TA_TENSOR_MEM_PROFILE 1
 
+/* Is TA::Tensor memory tracing enabled? */
+#cmakedefine TA_TENSOR_MEM_TRACE 1
+
 /* Is TTG available? */
 #cmakedefine TILEDARRAY_HAS_TTG 1
 

src/TiledArray/range.h

@@ -591,7 +591,9 @@ class Range {
 
   /// Move Constructor
 
-  /// \param other The range to be copied
+  /// \param[in,out] other The range to be copied; set to default (null) state
+  ///                on return
+  /// \post `other == Range{}`
   Range(Range_&& other)
       : datavec_(std::move(other.datavec_)),
         offset_(other.offset_),
@@ -639,7 +641,8 @@ class Range {
 
   /// Move assignment operator
 
-  /// \param other The range to be copied
+  /// \param[in,out] other The range to be copied; set to default (null) state
+  ///                on return
   /// \return A reference to this object
   /// \throw nothing
   Range_& operator=(Range_&& other) {