ARM64 Windows 10 IoT工控主板运行x86程序效率测试

        ARM上的 Windows 10 IoT 企业版支持仿真 x86 应用程序，而 ARM上的 Windows 11 IoT 企业版则支持仿真 x86 和 x64 应用程序。英创推出的名片尺寸ARM64工控主板ESM8400，可预装正版Windows 10 IoT企业版操作系统，x86程序可无需修改而直接在ESM8400上运行。

        下会将编写一个小程序，分别构建成x86和ARM64格式来测试其运行效率。所设计的测试程序代码如下，其中的TestSmp函数有两个输入参数，第一参数表示要创建测试线程的数量，第二个参数为所创建线程的运行时长。cbTestSmp是被创建的测试线程，测试线程主要是在一个while循环中，反复读取内存变量然后与预设值进行比较，在运行设定的时间后自动退出循环，其中的threadParam->loops变量会记录下while循环总共执行的次数。

typedef struct _SMP_THREAD_PARAM {     UINT32 durationMs;     UINT32 cpuId;     UINT64 loops;     BOOL   bSetAffinity;     UINT32 sandBoxSize;     LPVOID sandBoxStart; }SMP_THREAD_PARAM, * PSMP_THREAD_PARAM; DWORD WINAPI cbTestSmp(LPVOID param) {     PSMP_THREAD_PARAM threadParam = (PSMP_THREAD_PARAM)param;     DWORD tStart = GetTickCount();     UINT8* buffer = (UINT8*)threadParam->sandBoxStart;     wprintf(L"Ahou, Thread %d, running for %d ms\r\n", threadParam->cpuId, threadParam->durationMs);     // Write to sandbox     for (UINT32 i = 0; i < threadParam->sandBoxSize; i++)     {         buffer[i] = (UINT8)(i);// * (UINT32)threadParam->loops);     }     while ((GetTickCount() - tStart) < threadParam->durationMs)     {         // Read back from sandbox         for (UINT32 i = 0; i < threadParam->sandBoxSize; i++)         {             //if (buffer[i] != (UINT8)(i * (UINT32)threadParam->loops) )             if (buffer[i] != (UINT8)(i))// * (UINT32)threadParam->loops) )             {                 wprintf(L"Thread %d : error at byte %d for loop %I64d !!\r\n",                     threadParam->cpuId, i, threadParam->loops);             }         }         threadParam->loops++;     }     wprintf(L"Thread %d : terminating\r\n", threadParam->cpuId);     return 0; } void TestSmp(UINT32 nCpus, UINT32 durationMs) {     UINT32 i;     PSMP_THREAD_PARAM threadParams;     HANDLE* threadHandles;     UINT64 totalLoops = 0;     UINT32 sandBoxSize = 1024 * 128; // 128 kB     HANDLE h_array[1];     threadParams = (PSMP_THREAD_PARAM)malloc(nCpus * sizeof(SMP_THREAD_PARAM));     if (threadParams == NULL)     {         wprintf(L"Failed allocating thread params !\r\n");         return;     }     threadHandles = (HANDLE*)malloc(nCpus * sizeof(HANDLE));     if (threadHandles == NULL)     {         wprintf(L"Failed allocating thread handles !\r\n");         return;     }     for (i = 0; i < nCpus; i++)     {         threadParams[i].bSetAffinity = TRUE;         threadParams[i].cpuId = i;         threadParams[i].durationMs = durationMs;         threadParams[i].loops = 0;         threadParams[i].sandBoxSize = sandBoxSize;         threadParams[i].sandBoxStart = malloc(sandBoxSize);         threadHandles[i] = CreateThread(NULL, 0, cbTestSmp, &threadParams[i], 0, NULL);         wprintf(L"Thread handle %d : 0x%x\r\n", i, threadHandles[i]);     }     h_array[0] = threadHandles[0];     DWORD res = WaitForSingleObject(h_array[0], INFINITE);     Sleep(500);     if (res == WAIT_TIMEOUT)     {         wprintf(L"Timeout waiting for threads !\r\n");     }     else     {         wprintf(L"All threads exited\r\n");     }     for (i = 0; i < nCpus; i++)     {         wprintf(L"Thread %d did run %I64d loops\r\n", i, threadParams[i].loops);         totalLoops += threadParams[i].loops;         free(threadParams[i].sandBoxStart);         CloseHandle(threadHandles[i]);     }     wprintf(L"Total number of loops %I64d (%I64d millions)\r\n", totalLoops, totalLoops / 1000000);     free(threadHandles);     free(threadParams); }

        将上述代码分别编译构建成x86格式和ARM64模式，设置while循环执行10000ms，在ESM8400上的测试结果如下：

ESM8400 Win10 ARM工控主板运行x86和ARM64程序效率对比

        可以看到相同的代码，构建成本机ARM64格式的运行效率是x86格式的2.2倍以上。

        基于微软系统以及其开发工具良好的兼容性，我很容易做了另一个对比实验，将上述代码不经修改直接在VS2008中编译成WEC7应用程序，在英创的几款WEC7工控主板上做了同样的测试，测试结果如下：

        ESM3354是英创10年前推出的第一款预装WEC7的工控主板，主CPU采用了TI的单核Cortex-A8芯片——AM3354，ESM3354目前仍在批量供货。而安装Windows 10 IoT的ESM8400工控主板，主CPU为NXP的i.MX8M Plus四核Cortex-A53，与10年前的ESM3354相比，ESM8400的性能有超过10倍的提升。

         ARM上的 Windows IoT 企业版可以让习惯使用 x86/x64 的设备开发人员快速进行软件开发，大多数适用于 Windows IoT 企业版的文档都适用于 ARM64 和 x86/x64。通过仿真技术，ARM上的 Windows IoT可按原样运行x86/x64程序而无需修改，而直接构建本机ARM64应用程序能获得最佳的性能、响应能力和能耗。