利用双核机制优化素数生成

以下代码在基于如下环境中编写和测试的：

编译器：Apple LLVM2.0（兼容GNU99），Objective-C

操作系统：Mac OS X Snow Leopard

生成程序：64位

测试环境：Mac Mini，CPU：2.00GHz，Intel Core2 Duo P7350； 内存：1GB DDR3；L2 Cache：3MB

 

 

 

 

// // main.m // math_test // // Created by Zenny Chen on 5/18/11. // Copyright 2011 GreenGames Studio. All rights reserved. // #import <Foundation/Foundation.h> extern unsigned long realtime_counter(void); #define MAX_PRIME_COUNT 30 * 1024 static unsigned primeBuffer[MAX_PRIME_COUNT]; static unsigned primeBufferNaive[MAX_PRIME_COUNT]; static void producePrimeNumber(void) { primeBufferNaive[0] = 2; unsigned primeIndex = 1; unsigned number = 3; // normal calculation while(primeIndex < MAX_PRIME_COUNT) { _Bool isNotPrime = 0; for(unsigned i = 0; i < primeIndex; i++) { if(number % primeBufferNaive[i] == 0) { isNotPrime = 1; break; } } if(!isNotPrime) { primeBufferNaive[primeIndex++] = number; } number += 2; } } static void producePrimeNumberOpt(void) { primeBuffer[0] = 2; unsigned primeIndex = 1; unsigned number = 3; // normal calculation while(primeIndex < 100) { _Bool isNotPrime = 0; for(unsigned i = 0; i < primeIndex; i++) { if(number % primeBuffer[i] == 0) { isNotPrime = 1; break; } } if(!isNotPrime) { primeBuffer[primeIndex++] = number; } number += 2; } __block volatile int isFinished = 0; __block volatile int isNotPrime = 0; __block volatile int primeIndexForBlockFrom; __block volatile int primeIndexForBlockTo; __block volatile int ifStart = 0; __block volatile unsigned blockNumber; dispatch_queue_t queue = dispatch_queue_create("zenny_chen_firstQueue", NULL); dispatch_async(queue, ^(void) { while(!isFinished) { if(ifStart) { const unsigned to = primeIndexForBlockTo; const unsigned number = blockNumber; for(unsigned i = primeIndexForBlockFrom; i < to; i++) { if(number % primeBuffer[i] == 0) { isNotPrime = 1; break; } if(isNotPrime == 1) break; } ifStart = 0; } else asm("pause"); } }); while(primeIndex < MAX_PRIME_COUNT) { isNotPrime = 0; primeIndexForBlockFrom = primeIndex / 2; primeIndexForBlockTo = primeIndex; blockNumber = number; ifStart = 1; for(unsigned i = 0; i < primeIndex / 2; i++) { if(number % primeBuffer[i] == 0) { isNotPrime = 1; break; } if(isNotPrime == 1) break; } while(ifStart != 0) asm("pause"); if(isNotPrime == 0) { primeBuffer[primeIndex++] = number; } number += 2; } isFinished = 1; } static void test(void) { unsigned long beginTicks, endTicks; memset(primeBufferNaive, 0, sizeof(primeBufferNaive)); beginTicks = realtime_counter(); producePrimeNumber(); endTicks = realtime_counter(); printf("The value is: %u/n", primeBufferNaive[0] + primeBufferNaive[5] + primeBufferNaive[9]); printf("The ticks: %lu/n", endTicks - beginTicks); memset(primeBuffer, 0, sizeof(primeBuffer)); beginTicks = realtime_counter(); producePrimeNumberOpt(); endTicks = realtime_counter(); printf("The value is: %u/n", primeBuffer[0] + primeBuffer[5] + primeBuffer[9]); printf("The ticks: %lu/n", endTicks - beginTicks); for(int i = 0; i < MAX_PRIME_COUNT; i++) { if(primeBuffer[i] != primeBufferNaive[i]) { printf("Not equal: @%d/n", i); return; } } puts("Equal!"); } int main (int argc, const char * argv[]) { NSAutoreleasePool * pool = [[NSAutoreleasePool alloc] init]; // insert code here... test(); [pool drain]; return 0; }

 

测试结果：非优化版平均输出周期数为4778049698，约2.4秒

               优化版本平均输出周期数为2510125320，约1.3秒

 

 

下面是测试代码中对unsigned long realtime_counter(void)函数的定义：

 

// // hi.s // math_test // // Created by Zenny Chen on 5/18/11. // Copyright 2011 GreenGames Studio. All rights reserved. // .text .align 2 .globl _realtime_counter _realtime_counter: rdtsc shl $32, %rdx or %rdx, %rax ret