Merged PR 10383: Align items while saving at 256-byte boundary

Align items while saving at 256-byte boundary. Maintains binary-compatibility.

CHANGELOG.md

@@ -22,6 +22,9 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 - Gradient-checkpointing
 
 ### Fixed
+- Make sure that items are 256-byte aligned during saving
+- Make explicit matmul functions respect setting of cublasMathMode
+- Fix memory mapping for mixed paramter models
 - Removed naked pointer and potential memory-leak from file_stream.{cpp,h}
 - Compilation for GCC >= 7 due to exception thrown in destructor
 - Sort parameters by lexicographical order during allocation to ensure consistent 

VERSION

@@ -1 +1 @@
-v1.8.15
+v1.8.18

src/common/binary.cpp

@@ -123,8 +123,7 @@ void saveItems(const std::string& fileName,
     headers.push_back(Header{item.name.size() + 1,
                              (size_t)item.type,
                              item.shape.size(),
-                             item.size()}); // item size without padding 
-                                            // @TODO: should this be done with padding as asked below?
+                             item.bytes.size()}); // binary item size with padding, will be 256-byte-aligned
   }
 
   size_t headerSize = headers.size();
@@ -152,10 +151,10 @@ void saveItems(const std::string& fileName,
 
   // Write out all values
   for(const auto& item : items)
-    pos += out.write(item.data(), item.size()); // writes out data without padding, not aligned, @BUGBUG?
-
-    // @TODO: find out if padding should be enforced for memory-mapped storage like this:
-    // pos += out.write(item.data(), item.bytes.size()); // writes out data with padding
+    pos += out.write(item.data(), item.bytes.size()); // writes out data with padding, keeps 256-byte boundary. 
+                                                      // Amazingly this is binary-compatible with V1 and aligned and 
+                                                      // non-aligned models can be read with the same procedure.
+                                                      // No version-bump required. Gets 5-8% of speed back when mmapped.
 }
 
 }  // namespace binary

src/graph/expression_graph.h

@@ -522,7 +522,9 @@ public:
     LOG(info, "Memory mapping model at {}", ptr);
     auto items = io::mmapItems(ptr);
     
-    // deal with default parameter object that might not be a mapped object
+    // Deal with default parameter set object that might not be a mapped object.
+    // This gets assigned during ExpressionGraph::setDevice(...) and by default 
+    // would contain allocated tensors. Here we replace it with a mmapped version.
     auto it = paramsByElementType_.find(defaultElementType_);
     if(it != paramsByElementType_.end()) {
       // there is parameter object for that type

src/translator/translator.h

@@ -73,7 +73,7 @@ public:
     for(auto model : models) {
       marian::filesystem::Path modelPath(model);
       ABORT_IF(modelPath.extension() != marian::filesystem::Path(".bin"), 
-              "Non-binarized models cannot be mapped");
+              "Non-binarized models cannot be mmapped");
       mmaps_.push_back(std::move(mio::mmap_source(model)));
     }
 #endif