matlab convn function GPU slower than CPU?

Question

I test the convn on two GPU: Quadro 6000 and Titan both take longer time than cpu.

A quick test can be done in matlab:

maxloop=1000;
for i=1:maxloop
   output2= convn(rand(320,1), rand([6,1,300]),'full');
end

for i=1:maxloop    
    goutput2= convn(gpuArray.rand(320,1),gpuArray.rand([6,1,300]), 'full');
end

It takes, 0.52s on CPU, but 7s on Quadro 6000 and 15s+- on Titan.

What I had tested:

1) If change the rand input to fixed, predefined values does not give any improvement.

2) Predefine GPU output(goutput2) doesn't help so much.

Quadro

Quadro 6000

Titan

I do run the same test as the first answer:

Matlab Convolution using gpu

Same result obtained when m=1000; n=100; k=5;

Elapsed time is 2.367453 seconds. %%%%GPU

Elapsed time is 27.502952 seconds. %%%%CPU

My question is what and why my own test code is running slower on GPU?

Reading the answer that you link to, your relatively small data size does not work very well on the GPU (probably due to setup cost for each run). — Eric J.
Could you try with more reliable benchmarking techniques as discussed in this solution. and here - Measure and Improve GPU Performance. — Divakar

Steven Du Steven Du · Accepted Answer · 2014-08-15T08:59:16

It looks like question Matlab Convolution using gpu at first glances. But after increase data size from [6,1,300] to [1000,1,1000], there is no improvement in the GPU loop. So it's not about "size" of data.

When I reshape data from [1000,1,1000] to [1000,1000], GPU runs faster and CPU runs slower than previous test. Code and timing are listed below:

clear all;

maxloop=1000;

r5=rand(5,1);
r1000=rand([1000,1,1000]); %3d array

for i=1:maxloop   
   cpu_output1= convn(r5, r1000,'full');  %3d cpu array
end

r1000=reshape(r1000,[1000,1000]);
for i=1:maxloop
   cpu_output2= convn(r5, r1000,'full'); %2d cpu array
end

gr5=gpuArray.rand(5,1);
gr1000=gpuArray.rand([1000,1,1000]);

for i=1:maxloop    
    gpu_output1= convn(gr5,gr1000, 'full'); %3d gpu array
end

gr3_1000=reshape(gr1000,[1000,1000]);
for i=1:maxloop    
    gpu_output2= convn(gr5,gr3_1000, 'full'); %2d gpu array
end

enter image description here

matlab convn function GPU slower than CPU?

1 Answers