GPU enabled functions react slow on first call

Dear community, I want to utilize the GPU of my system to accelerate matrix multiplication. From the list of GPU-enabled functions, I can see that "mtimes" is available, therefore it should be accessible using gpuArrays.

My system is:

Matlab R2016b

Visual Studio 2013

CUDA 9.0

GTX 1070

I start by defining two matrices as gpuArrays:

 >>gDexMat=gpuArray(DexMat);
>> gx_PD=gpuArray(x_PD);
>> whos gDexMat gx_PD
  Name              Size                Bytes  Class       Attributes

gDexMat      200000x1000                  4  gpuArray              
gx_PD          1000x200000                4  gpuArray

I then perform the multiplication, gather the data and stop the time how long it took:

 >> tic;gIQMat=gx_PD*gDexMat;IQMat=gather(gIQMat);toc %doing the demodulation
Elapsed time is 0.528126 seconds.
>> tic;gIQMat=gx_PD*gDexMat;IQMat=gather(gIQMat);toc
Elapsed time is 0.414626 seconds.
>> tic;gIQMat=gx_PD*gDexMat;IQMat=gather(gIQMat);toc
Elapsed time is 0.405255 seconds.

In this example, the time it took is pretty constant. However, when I start Matlab the first time, the execution may take extremely long, up to 45 seconds. Even the first call:

 >>gDexMat=gpuArray(DexMat);

Can sometimes take up to a minute.

Does any body have any ideas why this is happening? Best, Ludwig

Answer 1

[Joss Knight]

MATLAB R2016b is built with CUDA 7.5. The first time you load the GPU libs they have to be just-in-time compiled to match your Pascal-architecture GPU. See the answer here for an explanation and some steps to prevent this compilation happening every time.

Answer 2

[scanner]

Thanks for your help. Installing Matlab R2017b solved this problem.