我尝试在Clojure中尽可能快地对复数数组进行乘法运算。所选数据结构是两个元素的映射，
这段Clojure代码的运行时间比等效的Java代码多30％（不包括启动时间）。这个函数能否重写为Clojure，以使其更快？Clojure代码中需要优化的函数是“multiply-complex-arrays”。
:re和:im，每个元素都是基本的Java本机双精度数组，以实现低内存开销。根据http://clojure.org/reference/java_interop，我使用了基本类型数组的确切类型规范。有了这些提示，aget被转换为本机数组dload操作，但存在两个效率问题，即循环计数器不是int而是long，因此每次索引数组时，计数器都会使用调用clojure/lang/RT.intCast将其转换为int。而aset也没有转换为本机操作，而是转换为调用clojure/lang/RT.aset。还有一个效率问题是checkcast。它检查每个循环，以确保数组实际上是双精度数组。这段Clojure代码的运行时间比等效的Java代码多30％（不包括启动时间）。这个函数能否重写为Clojure，以使其更快？Clojure代码中需要优化的函数是“multiply-complex-arrays”。
(def size 65536)
(defn get-zero-complex-array
[]
{:re (double-array size)
:im (double-array size)})
(defn multiply-complex-arrays
[a b]
(let [
a-re-array (doubles (get a :re))
a-im-array (doubles (get a :im))
b-re-array (doubles (get b :re))
b-im-array (doubles (get b :im))
res-re-array (double-array size)
res-im-array (double-array size)
]
(loop [i (int 0) size (int size)]
(if (< i size)
(let [
a-re (aget a-re-array i)
a-im (aget a-im-array i)
b-re (aget b-re-array i)
b-im (aget b-im-array i)
]
(aset res-re-array i (- (* a-re b-re) (* a-im b-im)))
(aset res-im-array i (+ (* a-re b-im) (* b-re a-im)))
(recur (unchecked-inc i) size))
{:re res-re-array :im res-im-array}))))
(let [
res (loop [i (int 0) a (get-zero-complex-array)]
(if (< i 30000)
(recur (inc i) (multiply-complex-arrays a a))
a))
]
(println (aget (get res :re) 0)))
multiply-complex-arrays 的主循环生成的 Java 汇编代码如下：
91: lload 8
93: lload 10
95: lcmp
96: ifge 216
99: aload_2
100: checkcast #51 // class "[D"
103: lload 8
105: invokestatic #46 // Method clojure/lang/RT.intCast:(J)I
108: daload
109: dstore 12
111: aload_3
112: checkcast #51 // class "[D"
115: lload 8
117: invokestatic #46 // Method clojure/lang/RT.intCast:(J)I
120: daload
121: dstore 14
123: aload 4
125: checkcast #51 // class "[D"
128: lload 8
130: invokestatic #46 // Method clojure/lang/RT.intCast:(J)I
133: daload
134: dstore 16
136: aload 5
138: checkcast #51 // class "[D"
141: lload 8
143: invokestatic #46 // Method clojure/lang/RT.intCast:(J)I
146: daload
147: dstore 18
149: aload 6
151: checkcast #51 // class "[D"
154: lload 8
156: invokestatic #46 // Method clojure/lang/RT.intCast:(J)I
159: dload 12
161: dload 16
163: dmul
164: dload 14
166: dload 18
168: dmul
169: dsub
170: invokestatic #55 // Method clojure/lang/RT.aset:([DID)D
173: pop2
174: aload 7
176: checkcast #51 // class "[D"
179: lload 8
181: invokestatic #46 // Method clojure/lang/RT.intCast:(J)I
184: dload 12
186: dload 18
188: dmul
189: dload 16
191: dload 14
193: dmul
194: dadd
195: invokestatic #55 // Method clojure/lang/RT.aset:([DID)D
198: pop2
199: lload 8
201: lconst_1
202: ladd
203: lload 10
205: lstore 10
207: lstore 8
209: goto 91
Java 代码：
class ComplexArray {
static final int SIZE = 1 << 16;
double re[];
double im[];
ComplexArray(double re[], double im[]) {
this.re = re;
this.im = im;
}
static ComplexArray getZero() {
return new ComplexArray(new double[SIZE], new double[SIZE]);
}
ComplexArray multiply(ComplexArray second) {
double resultRe[] = new double[SIZE];
double resultIm[] = new double[SIZE];
for (int i = 0; i < SIZE; i++) {
double aRe = this.re[i];
double aIm = this.im[i];
double bRe = second.re[i];
double bIm = second.im[i];
resultRe[i] = aRe * bRe - aIm * bIm;
resultIm[i] = aRe * bIm + bRe * aIm;
}
return new ComplexArray(resultRe, resultIm);
}
public static void main(String args[]) {
ComplexArray a = getZero();
for (int i = 0; i < 30000; i++) {
a = a.multiply(a);
}
System.out.println(a.re[0]);
}
}
在Java代码中组装相同的循环：
13: iload 4
15: ldc #5 // int 65536
17: if_icmpge 92
20: aload_0
21: getfield #2 // Field re:[D
24: iload 4
26: daload
27: dstore 5
29: aload_0
30: getfield #3 // Field im:[D
33: iload 4
35: daload
36: dstore 7
38: aload_1
39: getfield #2 // Field re:[D
42: iload 4
44: daload
45: dstore 9
47: aload_1
48: getfield #3 // Field im:[D
51: iload 4
53: daload
54: dstore 11
56: aload_2
57: iload 4
59: dload 5
61: dload 9
63: dmul
64: dload 7
66: dload 11
68: dmul
69: dsub
70: dastore
71: aload_3
72: iload 4
74: dload 5
76: dload 11
78: dmul
79: dload 9
81: dload 7
83: dmul
84: dadd
85: dastore
86: iinc 4, 1
89: goto 13
lein new app tmp2. 编辑 tmp/src/tmp/code.clj - 将def size、defn get-zero-complex-array和defn multiply-complex-arrays放置在(:gen-class))和(defn -main之间。3. 运行lein uberjar。4. 从 target/uberjar/tmp-0.1.0-SNAPSHOT.jar 中提取 tmp/core$multiply_complex_arrays.class。5. 执行javap -p -c core$multiply_complex_arrays >src。 - Fyodor Menshikov(set! *unchecked-math* true)这行代码将会把intCast的调用转换成l2i指令。 - sw1nn