Trading package size for float speed

Author: dcodeIO

.eslintrc.json

@@ -5,6 +5,9 @@
     },
     "globals": {
         "Uint8Array": true,
+        "Float32Array": true,
+        "Float64Array": true,
+        "define": true,
         "XMLHttpRequest": true,
         "Promise": true
     },

bench/bench.json

@@ -13,7 +13,9 @@
             "sint32": -42
         },
         "outer" : {
-            "bool" : [ true, false, false, true, false, false, true ]
+            "bool" : [ true, false, false, true, false, false, true ],
+            "double": 0.5
         }
-    }
+    },
+    "float": 0.25
 }

bench/bench.proto

@@ -5,6 +5,7 @@ message Test {
     string  string = 1;
     uint32  uint32 = 2;
     Inner   inner  = 3;
+    float   float  = 4;
 
     message Inner {
 
@@ -33,4 +34,5 @@ message Test {
 message Outer {
 
     repeated bool bool = 1;
+    double double = 2;
 }

bench/write.js

@@ -1,45 +1,75 @@
 var protobuf = require("../src/index"),
     newSuite = require("./suite");
 
+// This benchmark compares raw data type performance of Uint8Array and Buffer.
+// This uses internal finish machinery because overall difference is otherwise mainly caused by allocation.
+
+var array  = new Uint8Array(8);
+var buffer = new Buffer(8);
+
+// The following is a possible alternative to float / double writing - where supported. Quite a bit faster.
+
+var f32 = new Float32Array(1);
+var i8 = new Uint8Array(f32.buffer);
+
+function writeF32Array(buf, pos, val) {
+    f32[0] = val;
+    for (var i = 0; i < 4; ++i)
+        buf[pos + i] = i8[i];
+}
+
+// raw float write speed
 newSuite("float")
 .add("Writer#float", function() {
     var writer = new protobuf.Writer();
     writer.float(0.1);
-    writer.finish();
+    writer._finish(writer.head, array);
+})
+.add("Writer#writeF32Array", function() {
+    var writer = new protobuf.Writer();
+    writer.push(writeF32Array, 4, 0.1);
+    writer._finish(writer.head, array);
 })
 .add("BufferWriter#float", function() {
     var writer = new protobuf.BufferWriter();
     writer.float(0.1);
-    writer.finish();
+    writer._finish(writer.head, buffer);
+})
+.add("BufferWriter#writeF32Array", function() {
+    var writer = new protobuf.BufferWriter();
+    writer.push(writeF32Array, 4, 0.1);
+    writer._finish(writer.head, buffer);
 })
 .run();
 
+// raw double write speed
 newSuite("double")
 .add("Writer#double", function() {
     var writer = new protobuf.Writer();
     writer.double(0.1);
-    writer.finish();
+    writer._finish(writer.head, array);
 })
 .add("BufferWriter#double", function() {
     var writer = new protobuf.BufferWriter();
     writer.double(0.1);
-    writer.finish();
+    writer._finish(writer.head, buffer);
 })
 .run();
 
-var bytes = [0, 0, 0, 0, 0, 0, 0, 0];
-var arrayBytes = new Uint8Array(bytes);
-var bufferBytes = Buffer.from(bytes);
+var arrayPlus1 = new Uint8Array(array.length + 1);
+var bufferPlus1 = new Buffer(buffer.length + 1);
 
+// raw bytes write speed
+// interestingly, Uint8Array#set is faster than Buffer#copy, but only on actual Uint8Arrays.
 newSuite("bytes")
 .add("Writer#bytes", function() {
     var writer = new protobuf.Writer();
-    writer.bytes(arrayBytes);
-    writer.finish();
+    writer.bytes(array);
+    writer._finish(writer.head, arrayPlus1);
 })
 .add("BufferWriter#bytes", function() {
     var writer = new protobuf.BufferWriter();
-    writer.bytes(bufferBytes);
-    writer.finish();
+    writer.bytes(buffer);
+    writer._finish(writer.head, bufferPlus1);
 })
 .run();