Merge pull request #242 from objectbox/fb-read-speedup

Read speedup

Author: vaind

benchmark/bin/native_pointers.dart

@@ -2,6 +2,7 @@ import 'dart:ffi';
 import 'dart:typed_data';
 
 import 'package:ffi/ffi.dart';
+import 'package:objectbox/src/native/bindings/nativemem.dart';
 import 'package:objectbox_benchmark/benchmark.dart';
 
 // Results (Dart SDK 2.12):
@@ -19,11 +20,13 @@ import 'package:objectbox_benchmark/benchmark.dart';
 // which is consistent with profiling objectbox-dart Box.read().
 
 void main() async {
-  final sizeInBytes = 256;
+  final sizeInBytes = 1024;
   await AsTypedList(sizeInBytes).report();
 
   // just checking if using a larger underlying type would help with anything
-  await AsTypedListUint64((sizeInBytes / 8).floor()).report();
+  // await AsTypedListUint64((sizeInBytes / 8).floor()).report();
+
+  await TypedListMemCopy(sizeInBytes).report();
 }
 
 class AsTypedList extends Benchmark {
@@ -73,3 +76,37 @@ class AsTypedListUint64 extends Benchmark {
     super.teardown();
   }
 }
+
+class TypedListMemCopy extends Benchmark {
+  final int length;
+  late final Pointer<Uint8> nativePtr;
+  late final Pointer<Uint8> nativePtr2;
+  late final ByteBuffer buffer;
+  late final ByteData data;
+
+  TypedListMemCopy(this.length)
+      : super('${TypedListMemCopy}', iterations: 1000);
+
+  @override
+  void runIteration(int i) {
+    memcpy(nativePtr, nativePtr2, length);
+    ByteData.view(buffer, length);
+    // actually using the data (read flatbuffers) doesn't matter here
+  }
+
+  @override
+  void setup() {
+    nativePtr = malloc<Uint8>(length);
+    nativePtr2 = malloc<Uint8>(length);
+    assert(nativePtr.asTypedList(length).offsetInBytes == 0);
+    buffer = nativePtr.asTypedList(length).buffer;
+    data = ByteData.view(buffer, 0);
+  }
+
+  @override
+  void teardown() {
+    malloc.free(nativePtr);
+    malloc.free(nativePtr2);
+    super.teardown();
+  }
+}

generator/lib/src/code_chunks.dart

@@ -412,8 +412,8 @@ class CodeChunks {
       }
     });
 
-    return '''(Store store, Uint8List fbData) {
-      final buffer = fb.BufferContext.fromBytes(fbData);
+    return '''(Store store, ByteData fbData) {
+      final buffer = fb.BufferContext(fbData);
       final rootOffset = buffer.derefObject(0);
       ${preLines.join('\n')}
       final object = ${entity.name}(${constructorLines.join(', \n')})${cascadeLines.join('\n')};

objectbox/lib/flatbuffers/flat_buffers.dart

@@ -44,10 +44,10 @@ class BufferContext {
   factory BufferContext.fromBytes(List<int> byteList) {
     Uint8List uint8List = _asUint8List(byteList);
     ByteData buf = new ByteData.view(uint8List.buffer, uint8List.offsetInBytes);
-    return new BufferContext._(buf);
+    return new BufferContext(buf);
   }
 
-  BufferContext._(this._buffer);
+  BufferContext(this._buffer);
 
   @pragma('vm:prefer-inline')
   int derefObject(int offset) {
@@ -883,6 +883,7 @@ class Float32Reader extends Reader<double> {
 
 class Int64Reader extends Reader<int> {
   const Int64Reader() : super();
+
   @override
   int get size => _sizeofInt64;
 

objectbox/lib/src/modelinfo/entity_definition.dart

@@ -13,7 +13,7 @@ import 'modelentity.dart';
 class EntityDefinition<T> {
   final ModelEntity model;
   final int Function(T, fb.Builder) objectToFB;
-  final T Function(Store, Uint8List) objectFromFB;
+  final T Function(Store, ByteData) objectFromFB;
   final int? Function(T) getId;
   final void Function(T, int) setId;
   final List<ToOne> Function(T) toOneRelations;

objectbox/lib/src/native/bindings/data_visitor.dart

@@ -36,6 +36,7 @@ Pointer<NativeFunction<obx_data_visitor>> dataVisitor(
 Pointer<NativeFunction<obx_data_visitor>> objectCollector<T>(
         List<T> list, Store store, EntityDefinition<T> entity) =>
     dataVisitor((Pointer<Uint8> data, int size) {
-      list.add(entity.objectFromFB(store, data.asTypedList(size)));
+      list.add(entity.objectFromFB(
+          store, InternalStoreAccess.reader(store).access(data, size)));
       return true;
     });

objectbox/lib/src/native/bindings/flatbuffers.dart

@@ -1,10 +1,10 @@
 import 'dart:ffi';
-import 'dart:io' show Platform;
 import 'dart:typed_data';
 
 import 'package:ffi/ffi.dart';
 
 import '../../../flatbuffers/flat_buffers.dart' as fb;
+import 'nativemem.dart';
 
 // ignore_for_file: public_member_api_docs
 
@@ -42,12 +42,6 @@ class BuilderWithCBuffer {
   Allocator get allocator => _allocator;
 }
 
-// FFI signature
-typedef _dart_memset = void Function(Pointer<Uint8>, int, int);
-typedef _c_memset = Void Function(Pointer<Uint8>, Int32, IntPtr);
-
-_dart_memset? fbMemset;
-
 class Allocator extends fb.Allocator {
   // We may, in practice, have only two active allocations: one used and one
   // for resizing. Therefore, we use a ring buffer of a fixed size (2).
@@ -98,34 +92,46 @@ class Allocator extends fb.Allocator {
   void clear(ByteData data, bool isFresh) {
     if (isFresh) return; // freshly allocated data is zero-ed out (see [calloc])
 
-    if (fbMemset == null) {
-      if (Platform.isWindows) {
-        try {
-          // DynamicLibrary.process() is not available on Windows, let's load a
-          // lib that defines 'memset()' it - should be mscvr100 or mscvrt DLL.
-          // mscvr100.dll is in the frequently installed MSVC Redistributable.
-          fbMemset = DynamicLibrary.open('msvcr100.dll')
-              .lookupFunction<_c_memset, _dart_memset>('memset');
-        } catch (_) {
-          // fall back if we can't load a native memset()
-          fbMemset = (Pointer<Uint8> ptr, int byte, int size) =>
-              ptr.cast<Uint8>().asTypedList(size).fillRange(0, size, byte);
-        }
-      } else {
-        fbMemset = DynamicLibrary.process()
-            .lookupFunction<_c_memset, _dart_memset>('memset');
-      }
-    }
-
     // only used for sanity checks:
     assert(_data[_index] == data);
     assert(_allocs[_index].address != 0);
 
-    fbMemset!(_allocs[_index], 0, data.lengthInBytes);
+    // TODO - there are other options to clear the builder, see how other
+    //        FlatBuffer implementations do it.
+    memset(_allocs[_index], 0, data.lengthInBytes);
   }
 
   void freeAll() {
     if (_allocs[0].address != 0) calloc.free(_allocs[0]);
     if (_allocs[1].address != 0) calloc.free(_allocs[1]);
   }
 }
+
+/// Implements a native data access wrapper to circumvent Pointer.asTypedList()
+/// slowness. The idea is to reuse the same buffer and rather memcpy the data,
+/// which ends up being faster than calling asTypedList(). Hopefully, we will
+/// be able to remove this if (when) asTypedList() gets optimized in Dart SDK.
+class ReaderWithCBuffer {
+  // See /benchmark/bin/native_pointers.dart for the max buffer size where it
+  // still makes sense to use memcpy. On Linux, memcpy starts to be slower at
+  // about 10-15 KiB. TODO test on other platforms to find an optimal limit.
+  static const _maxBuffer = 4 * 1024;
+  final _bufferPtr = malloc<Uint8>(_maxBuffer);
+  late final ByteBuffer _buffer = _bufferPtr.asTypedList(_maxBuffer).buffer;
+
+  ReaderWithCBuffer() {
+    assert(_bufferPtr.asTypedList(_maxBuffer).offsetInBytes == 0);
+  }
+
+  void clear() => malloc.free(_bufferPtr);
+
+  ByteData access(Pointer<Uint8> dataPtr, int size) {
+    if (size > _maxBuffer) {
+      final uint8List = dataPtr.asTypedList(size);
+      return ByteData.view(uint8List.buffer, uint8List.offsetInBytes, size);
+    } else {
+      memcpy(_bufferPtr, dataPtr, size);
+      return ByteData.view(_buffer, 0, size);
+    }
+  }
+}

objectbox/lib/src/native/bindings/helpers.dart

@@ -7,6 +7,7 @@ import '../../common.dart';
 import '../../modelinfo/entity_definition.dart';
 import '../store.dart';
 import 'bindings.dart';
+import 'flatbuffers.dart';
 
 // ignore_for_file: public_member_api_docs
 
@@ -87,9 +88,10 @@ class CursorHelper<T> {
   final EntityDefinition<T> _entity;
   final Store _store;
   final Pointer<OBX_cursor> ptr;
+  late final ReaderWithCBuffer _reader = InternalStoreAccess.reader(_store);
 
   final bool _isWrite;
-  late final Pointer<Pointer<Void>> dataPtrPtr;
+  late final Pointer<Pointer<Uint8>> dataPtrPtr;
 
   late final Pointer<IntPtr> sizePtr;
 
@@ -106,8 +108,7 @@ class CursorHelper<T> {
     }
   }
 
-  Uint8List get readData =>
-      dataPtrPtr.value.cast<Uint8>().asTypedList(sizePtr.value);
+  ByteData get readData => _reader.access(dataPtrPtr.value, sizePtr.value);
 
   EntityDefinition<T> get entity => _entity;
 
@@ -131,13 +132,13 @@ class CursorHelper<T> {
 }
 
 T withNativeBytes<T>(
-    Uint8List data, T Function(Pointer<Void> ptr, int size) fn) {
+    Uint8List data, T Function(Pointer<Uint8> ptr, int size) fn) {
   final size = data.length;
   assert(size == data.lengthInBytes);
   final ptr = malloc<Uint8>(size);
   try {
     ptr.asTypedList(size).setAll(0, data); // copies `data` to `ptr`
-    return fn(ptr.cast<Void>(), size);
+    return fn(ptr, size);
   } finally {
     malloc.free(ptr);
   }

objectbox/lib/src/native/bindings/nativemem.dart

@@ -0,0 +1,25 @@
+import 'dart:ffi';
+import 'dart:io';
+
+/// Provides native memory manipulation, operating on FFI Pointer<Void>.
+
+/// memset(ptr, value, num) sets the first num bytes of the block of memory
+/// pointed by ptr to the specified value (interpreted as an uint8).
+final _dart_memset memset =
+    _stdlib.lookupFunction<_c_memset, _dart_memset>('memset');
+
+/// memcpy (destination, source, num) copies the values of num bytes from the
+/// data pointed to by source to the memory block pointed to by destination.
+final _dart_memcpy memcpy =
+    _stdlib.lookupFunction<_c_memcpy, _dart_memcpy>('memcpy');
+
+// FFI signature
+typedef _dart_memset = void Function(Pointer<Uint8>, int, int);
+typedef _c_memset = Void Function(Pointer<Uint8>, Int32, IntPtr);
+
+typedef _dart_memcpy = void Function(Pointer<Uint8>, Pointer<Uint8>, int);
+typedef _c_memcpy = Void Function(Pointer<Uint8>, Pointer<Uint8>, IntPtr);
+
+final DynamicLibrary _stdlib = Platform.isWindows // no .process() on windows
+    ? DynamicLibrary.open('vcruntime140.dll') // required by objectbox.dll
+    : DynamicLibrary.process();