When compiling with cmake, build products also contain tools used for testing, as explained below.

MNNV2Basic.out

Usage

Test performance and dump tensor data.

Parameters

./MNNV2Basic.out temp.mnn 10 0 0 4 1x3x224x224

• The first parameter specifies the file name of MNN model.
• The second parameter specifies the number of loop times for the performance test, and 10 indicates that it will run 10 times to test the performance.
• The third parameter specifies whether to dump the the inference intermediate tensor. 0: don’t dump, 1: dump output of all operators, 2: dump the input and output for all operators. If dump, the directory is “output”.
• The fourth parameter specifies the computing device that performs the inference. Valid values are 0 (floating point CPU), 1 (Metal), 3 (OpenCL), 6 (OpenGL), 7 (Vulkan).
• The fifth parameter specifies the thread number, default set as 4, only valid for CPU.
• The sixth parameter specifies the size of the input tensor, which generally does not need to be specified.

Default input and output

Only single input, single output is supported. The input is read from “input_0.txt” and the output is dumped to “output.txt”.

checkFile.out

Usage

Check if the two tensor text files are consistent.

Parameters

./checkFile.out XXX.txt YYY.txt 0.1.

• 0.1 means absolute threshold, 0.0001 if not specified.
• When the comparison value exceeds the absolute threshold, it will be output directly to the console.

checkDir.out

Usage

Compare the files with the same name in the two folders.

Parameters

./checkDir.out output android_output 1

• 1 means absolute threshold, 0.0001 if not specified.
• When the comparison value exceeds the absolute threshold, it will be output directly to the console.

timeProfile.out

Usage

Operators’ total time-consuming statistic.

Parameters

./timeProfile.out temp.mnn 10 0 1x3x224x224

• The first parameter is filename of model.
• The second parameter is run times, default 100.
• The third parameter is the forward type, default 0.
• The fourth parameter is input tensor size, generally needn’t be specified.

Outputs

• The first column is the operator’s type.
• The second column is the average time consuming.
• The third column is time-consuming percent.
• Example:

  Node Type                Avg(ms)       %             Called times
  Softmax                  0.018100      0.022775      1.000000
  Pooling                  0.080800      0.101671      1.000000
  ConvolutionDepthwise     14.968399     18.834826     13.000000
  Convolution              64.404617     81.040726     15.000000
  total time : 79.471924 ms, total mflops : 2271.889404

backendTest.out

Usage

This tool compares the results of the inferences performed by the specified computing device and CPU. Read input_0.txt in the current directory as input by default.

Parameters

./backendTest.out temp.mnn 3 0.15

• First parameter: model file.
• Second parameter: computing device that performs inference.
• Third parameter: error tolerance.