windows下运行nvprof --metrics命令报错？cuda分析错误

Question

windows下运行nvprof --metrics命令报错\xef\xbc\x9a

\n\n

==6580== NVPROF is profiling process 6580, command: Project1.exe\n==6580== Error: Internal profiling error 4292:1.\n======== Error: CUDA profiling error.\n

\n\n

错误1

\n\n

如果我只使用nvprof命令，不会报错\xef\xbc\x9a

\n\n

F:\\vstest\\Project1\\x64\\Release>nvprof Project1.exe\n==384== NVPROF is profiling process 384, command: Project1.exe\nsumMatrixOnGPU2D <<<(512,512), (32,32)>>> elapsed 22 ms\n==384== Profiling application: Project1.exe\n==384== Profiling result:\n            Type  Time(%)      Time     Calls       Avg       Min       Max  Name\n GPU activities:   61.28%  538.11ms         2  269.06ms  260.98ms  277.13ms  [CUDA memcpy HtoD]\n                   36.29%  318.68ms         1  318.68ms  318.68ms  318.68ms  [CUDA memcpy DtoH]\n                    2.43%  21.364ms         1  21.364ms  21.364ms  21.364ms  sumMatrixOnGPU2D(float*, float*, float*, int, int)\n      API calls:   56.77%  1.29771s         3  432.57ms  47.895ms  1.19911s  cudaMalloc\n                   37.53%  857.94ms         3  285.98ms  261.20ms  319.19ms  cudaMemcpy\n                    2.56%  58.617ms         1  58.617ms  58.617ms  58.617ms  cudaDeviceReset\n                    2.13%  48.594ms         3  16.198ms  14.312ms  17.671ms  cudaFree\n                    0.95%  21.732ms         2  10.866ms  275.60us  21.456ms  cudaDeviceSynchronize\n                    0.02%  512.70us         1  512.70us  512.70us  512.70us  cudaLaunchKernel\n                    0.02%  359.30us        96  3.7420us     100ns  204.60us  cuDeviceGetAttribute\n                    0.02%  347.80us         1  347.80us  347.80us  347.80us  cudaGetDeviceProperties\n                    0.01%  180.60us         1  180.60us  180.60us  180.60us  cuDeviceGetPCIBusId\n                    0.00%  32.100us         1  32.100us  32.100us  32.100us  cuDeviceTotalMem\n                    0.00%  13.400us         1  13.400us  13.400us  13.400us  cudaSetDevice\n                    0.00%  4.0000us         3  1.3330us     200ns  3.5000us  cuDeviceGetCount\n                    0.00%  3.9000us         1  3.9000us  3.9000us  3.9000us  cudaGetLastError\n                    0.00%  1.1000us         2     550ns     200ns     900ns  cuDeviceGet\n                    0.00%  1.0000us         1  1.0000us  1.0000us  1.0000us  cuDeviceGetName\n                    0.00%     300ns         1     300ns     300ns     300ns  cuDeviceGetUuid\n                    0.00%     300ns         1     300ns     300ns     300ns  cuDeviceGetLuid\n

\n\n

我想问一下是什么问题，如何使用命令 nvprof --metrics

\n

Answer 1

我找到了答案。

我正在添加对我有用的解决方案，作为其他人的参考。

您需要打开 NVIDIA 控制面板（右键单击桌面，然后选择它）桌面（从顶部菜单）-> 启用开发人员设置。然后，选择开发人员（从侧树中）-> 管理 GPU 性能计数器 -> 允许所有用户访问 GPU 性能计数器。

NVidia 之前对此进行了详细记录；现在很难找到了。