performance_gpu test timing is wrong (Patch #1846)

In the opencv/samples/gpu/performance/ there is a program that relates gpu performance versus cpu performance. This is a very neat and handy tool, but, it has a shortcoming which i expected. A thing that is commonly stated is the problem of the uploading of the GPU image to the GPU memory. This is for the GPU an extra step, and should in my opinion therefore be taken in to account in the computation time.

Proposition¶

tests.cpp (current)

   d_templ.upload(templ);
   GPU_ON;
   gpu::matchTemplate(d_src, d_templ, d_dst, CV_TM_CCORR);
   GPU_OFF;

test.cpp (including uploading to mem)

   GPU_ON;
   d_templ.upload(templ);
   gpu::matchTemplate(d_src, d_templ, d_dst, CV_TM_CCORR);
   GPU_OFF;

Argumentation¶

I think this should be done, because you are comparing CPU and GPU with the consideration to switch from CPU to GPU, this does mean that you have to take the memcpy in account.
The difference is obviously significant for small files/functions, but small for large computations that take >>100ms. In this way, you can immediatelly see that small functions are way less efficient for the gpu than large functions+files. This fact is ignored by the tests.

Example:¶

Current method

minMaxLoc:
    14.5524   0.432069  *x33.7*         2000x2000, 32F

New proposition

minMaxLoc:
    14.7356   0.530901  *x27.8*         2000x2000, 32F