MKL VSL convolution memory issue

[anniezh]

Hello,

I used Intel MKL VSL functions to do image convolution. We found normally the VSL_CONV_MODE_FFT mode runs faster for large image and large kernel combinations. However, I tried an 8k by 8k 4byte float image and a 5 by 5 float kernel with VslsConv2d in VSL_CONV_MODE_FFT mode. The process took almost all my computer remaining memory about 3G and a few minutes to finish. If we try a 7k by 7k float image and a 5 by 5 kernel. It only takes a few seconds to finish. Is it a bug in MKL? Thanks.

[Gennady_F_Intel]

Hello, 

that's happens because of swapping problem and the RAM size is not enough on your system.

for example, I checked how it works on the my side ( RAM ==32 Gb) and here are the results I had:

size == 5000, execTime == 0.610678 sec

size == 7000, execTime == 0.730076 sec

size == 8000, execTime == 0.761053 sec

checked with the latest MKL v.10.3 Update 11 and MKL 11.0 ( which we are going to release soon)

OS - Lin64, lp64. threading 

--Gennady