gcc mthumb + O2 + mfast-math 选项测试

虽然arm cpu是32的，但是 sdram为16为总线，所以猜测 16 位的 thumb指令可以带来一些优化。

修改了原来的测试代码，避免 O2 把所有代码优化掉（这可是O2的特长）

测试结果表明，对于我们的浮点运算，这两个选项作用远没有 –ffast-math明显。

测试结果也表明， 对我们这个测试程序，–ffast-math 对精度没有影响。

测试方法

分别使用marvell（gcc 4.1.1 ） 以及 maemo 编译器（codesourcery 2005q3-2，gcc3.4.4）并使用不同的编译选项进行测试对比。 程序运行在 PXA310 624M CPU 下。

 

测试代码

#include <stdlib.h>   /* exit */

#include <sys/time.h> /* struct timeval and friends */

#include <time.h>     /* gettimeofday */

#define MAX_DIVIDEND 1000000.231

#define MIN_DIVIDEND 0.29

#define STEP_DIVIDEND 0.33

#define DIVISOR 23.0

#define BUFFER_SIZE 200

/**

 * Utility to return a pointer to a statically allocated buffer that

 * holds the text representation of seconds since this program was

 * started. Not safe to use in threaded programs!

 */

static void timestamp(const char* buffer) {

  /* Holds the starting timestamp. 0 means that it has not been

     initialized. */

  static int startSecond = 0;

  static int startMs = 0;

  /* Temp storage for the secs + microseconds time. */

  struct timeval tv;

  /* Temp storage for the difference between start and now. */

  int deltaSecond, deltaMs;

 

  /* Get current time and convert into microseconds flat. */

  gettimeofday(&tv, NULL);

  /* Running for the first time? */

  if (startSecond == 0) {

    /* Copy to prev so that we get 0 delta. */

    startSecond = tv.tv_sec;

    startMs = tv.tv_usec;

  }

 

  /* Calculate the delta (in microseconds). */

  deltaSecond = tv.tv_sec - startSecond;

  deltaMs = tv.tv_usec - startMs;

 

  /* Create the string giving offset from start in seconds. */

  snprintf(buffer, BUFFER_SIZE, "%u.%u",deltaSecond,deltaMs);

}

 

 

int main(int argc, char * argv[])

{

         double divident, result;

         double result1;

 

    char buffer[BUFFER_SIZE];

    char buffer1[BUFFER_SIZE];

    char buffer2[BUFFER_SIZE];

    char buffer3[BUFFER_SIZE];

    char buffer4[BUFFER_SIZE];

 

    timestamp(buffer);

         for(divident=MIN_DIVIDEND; divident<MAX_DIVIDEND; divident+=STEP_DIVIDEND) {

                   result = divident/DIVISOR;

                   result1 += result;

         }

 

         timestamp(buffer1);

         for(divident=MIN_DIVIDEND; divident<MAX_DIVIDEND; divident+=STEP_DIVIDEND) {

                   result = divident*DIVISOR;

                   result1 += result;

         }

 

         timestamp(buffer2);

         for(divident=MIN_DIVIDEND; divident<MAX_DIVIDEND; divident+=STEP_DIVIDEND) {

                   result = divident+DIVISOR;

                   result1 += result;

         }

 

         timestamp(buffer3);

         for(divident=MIN_DIVIDEND; divident<MAX_DIVIDEND; divident+=STEP_DIVIDEND) {

                   result = divident-DIVISOR;

                   result1 += result;

         }

         timestamp(buffer4);

 

    printf("Start time is: %s/n",buffer);

         printf("DIV End time is: %s/n",buffer1);

         printf("MUL End time is: %s/n",buffer2);

         printf("ADD End time is: %s/n",buffer3);

         printf("SUB End time is: %s/n",buffer4);

         printf("result1:%f/n", result1);

         return 0;

}

Marvell 编译器编译

arm-iwmmxt-linux-gnueabi-gcc float.c -o float1

arm-iwmmxt-linux-gnueabi-objdump -D float1 > float1d

arm-iwmmxt-linux-gnueabi-strip float1

 

arm-iwmmxt-linux-gnueabi-gcc -O2 float.c -o float2

arm-iwmmxt-linux-gnueabi-objdump -D float2 > float2d

arm-iwmmxt-linux-gnueabi-strip float2

 

arm-iwmmxt-linux-gnueabi-gcc -ffast-math float.c -o float3

arm-iwmmxt-linux-gnueabi-objdump -D float3 > float3d

arm-iwmmxt-linux-gnueabi-strip float3

 

arm-iwmmxt-linux-gnueabi-gcc -ffast-math -O2 float.c -o float4

arm-iwmmxt-linux-gnueabi-objdump -D float4 > float4d

arm-iwmmxt-linux-gnueabi-strip float4

 

arm-iwmmxt-linux-gnueabi-gcc -mthumb -ffast-math float.c -o float5

arm-iwmmxt-linux-gnueabi-objdump -D float5 > float5d

arm-iwmmxt-linux-gnueabi-strip float5

 

arm-iwmmxt-linux-gnueabi-gcc -mthumb -ffast-math -O2 float.c -o float6

arm-iwmmxt-linux-gnueabi-objdump -D float6 > float6d

arm-iwmmxt-linux-gnueabi-strip float6

测试结果

/mnt/share # ./float1;./float2;./float3;./float4;./float5;./float6

Start time is: 0.0

DIV End time is: 4.4294917314

MUL End time is: 5.360439

ADD End time is: 7.4294759123

SUB End time is: 8.257485

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 4.4294863363

MUL End time is: 5.353536

ADD End time is: 7.4294732719

SUB End time is: 8.177472

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294389481

MUL End time is: 3.4294799614

ADD End time is: 5.4294230977

SUB End time is: 6.4294696195

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294326050

MUL End time is: 3.4294682183

ADD End time is: 4.131570

SUB End time is: 6.4294510822

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294628850

MUL End time is: 3.315699

ADD End time is: 5.4294958414

SUB End time is: 7.4294668472

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 1.451045

MUL End time is: 2.895364

ADD End time is: 4.361311

SUB End time is: 5.861070

result1:37944672758190.882812

/mnt/share # ./float1;./float2;./float3;./float4;./float5;./float6

Start time is: 0.0

DIV End time is: 4.4294916422

MUL End time is: 5.359094

ADD End time is: 6.790939

SUB End time is: 8.256269

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 4.4294863825

MUL End time is: 5.353796

ADD End time is: 7.4294733205

SUB End time is: 8.178055

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294389534

MUL End time is: 3.4294799570

ADD End time is: 4.263373

SUB End time is: 6.4294696403

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 1.358483

MUL End time is: 3.4294682785

ADD End time is: 4.132290

SUB End time is: 6.4294511573

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 1.661561

MUL End time is: 3.315752

ADD End time is: 5.4294958877

SUB End time is: 6.700518

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294418996

MUL End time is: 3.4294863151

ADD End time is: 5.4294328820

SUB End time is: 6.4294828801

result1:37944672758190.882812

 

Meamo 编译器编译

[sbox-CHINOOK_ARMEL: ~] > cat 1.sh

gcc -mfloat-abi=soft -march=iwmmxt float.c -o cs_float1

objdump -D cs_float1 > cs_float1d

strip cs_float1

 

gcc -mfloat-abi=soft -march=iwmmxt -O2 float.c -o cs_float2

objdump -D cs_float2 > cs_float2d

strip cs_float2

 

gcc -mfloat-abi=soft -march=iwmmxt -ffast-math float.c -o cs_float3

objdump -D cs_float3 > cs_float3d

strip cs_float3

 

gcc -mfloat-abi=soft -march=iwmmxt -ffast-math -O2 float.c -o cs_float4

objdump -D cs_float4 > cs_float4d

strip cs_float4

 

gcc -mfloat-abi=soft -march=iwmmxt -mthumb -ffast-math float.c -o cs_float5

objdump -D cs_float5 > cs_float5d

strip cs_float5

 

gcc -mfloat-abi=soft -march=iwmmxt -mthumb -ffast-math -O2 float.c -o cs_float6

objdump -D cs_float6 > cs_float6d

strip cs_float6

测试结果

/mnt/share # ./cs_float1;./cs_float2;./cs_float3;./cs_float4;./cs_float5;./cs_float6

Start time is: 0.0

DIV End time is: 5.4294362943

MUL End time is: 7.4294184865

ADD End time is: 9.4294103937

SUB End time is: 11.4294057006

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 4.371613

MUL End time is: 6.208795

ADD End time is: 8.142781

SUB End time is: 10.145741

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294797294

MUL End time is: 4.4294618808

ADD End time is: 5.610039

SUB End time is: 7.564100

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 1.843834

MUL End time is: 3.641867

ADD End time is: 5.536020

SUB End time is: 7.503841

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294952400

MUL End time is: 4.4294930974

ADD End time is: 6.38597

SUB End time is: 8.148249

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294830987

MUL End time is: 4.4294686170

ADD End time is: 6.4294639970

SUB End time is: 8.4294627964

result1:37944672758190.882812

/mnt/share # ./cs_float1;./cs_float2;./cs_float3;./cs_float4;./cs_float5;./cs_float6

Start time is: 0.0

DIV End time is: 4.396515

MUL End time is: 6.217185

ADD End time is: 8.137225

SUB End time is: 10.91183

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 4.371910

MUL End time is: 6.208795

ADD End time is: 8.143338

SUB End time is: 10.146496

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294796584

MUL End time is: 4.4294619369

ADD End time is: 6.4294578720

SUB End time is: 8.4294533344

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294811396

MUL End time is: 4.4294608746

ADD End time is: 6.4294504581

SUB End time is: 7.505244

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294953577

MUL End time is: 4.4294932613

ADD End time is: 6.41299

SUB End time is: 8.152282

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294830429

MUL End time is: 3.719345

ADD End time is: 5.673306

SUB End time is: 7.661928

result1:37944672758190.882812

测试结果（运行在meamo文件系统）

Nokia-N810-42-19:/mnt/share# ./cs_float1;./cs_float2;./cs_float3;./cs_float4;./cs_float5;./cs_float6

Start time is: 0.0

DIV End time is: 4.376203

MUL End time is: 6.189725

ADD End time is: 8.100677

SUB End time is: 10.45632

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 4.350762

MUL End time is: 6.178718

ADD End time is: 8.104297

SUB End time is: 10.97945

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294787951

MUL End time is: 3.634264

ADD End time is: 5.584049

SUB End time is: 7.528977

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294802638

MUL End time is: 4.4294591854

ADD End time is: 6.4294478357

SUB End time is: 8.4294437883

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294943224

MUL End time is: 4.4294912370

ADD End time is: 6.11563

SUB End time is: 8.111958

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294821997

MUL End time is: 4.4294669333

ADD End time is: 6.4294614187

SUB End time is: 8.4294593228

result1:37944672758190.882812

Nokia-N810-42-19:/mnt/share# ./cs_float1;./cs_float2;./cs_float3;./cs_float4;./cs_float5;./cs_float6

Start time is: 0.0

DIV End time is: 4.377758

MUL End time is: 6.191395

ADD End time is: 8.102222

SUB End time is: 10.47143

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 4.351143

MUL End time is: 6.179261

ADD End time is: 8.104836

SUB End time is: 10.98367

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294787966

MUL End time is: 3.634115

ADD End time is: 5.584012

SUB End time is: 7.528855

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294802441

MUL End time is: 4.4294591819

ADD End time is: 6.4294478172

SUB End time is: 8.4294438132

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294943712

MUL End time is: 4.4294912686

ADD End time is: 6.11965

SUB End time is: 8.112546

result1:37944672758190.882812

Start time is: 0.0

DIV End time is: 2.4294821782

MUL End time is: 4.4294669237

ADD End time is: 5.646800

SUB End time is: 7.626067

result1:37944672758190.882812