Tablet shield versus jetson k1

Hello,

I program on both shield tablet k1 and jetson k1. I found somethink strange about the processing duration between shield tablet and jetson k1.

For the same program between android and ubuntu there is a huge difference. For example a simple sobel image processing takes about 2ms but on the jetson it takes 8ms.

Why there are so huge difference between these architecture?

Hi xav123,

Would you please provide more information of you Shield tablet and Jetson TK1? Such as BSP/SW version info, installed libraries and toolkits…etc.

Please also provide your problem that can run on both android and ubuntu system as reference.

Besides, could you confirm whether your Shield tablet is running on “Performance mode”, and try to set Jetson in max performance mode also.

Thanks

Hello,

Sorry for the delay.

I just add the programs I use on Android and Jetson on my github.

https://github.com/xav12358/SobelK1

For the version of my Jetson, it is a trusty version : (from the uname -a command)
Linux tegra-ubuntu 3.10.40-gdacac96 #1 SMP PPREEMPT Thu Jun 25 15:25:11 PDT armv7l armv7l armv7l GNU/Linux

I use the 6.5 cuda-toolkit on the Jetson and on the Shield.

I try to put the “Performance mode” on my Jetson but it doesn’t change the process’s duration.

If you need more information ask me.

I just add the duration on each platform on the github folder.

I think I forget to say that the sobel processing is made on cuda kernel.
On the shield tablet the sobel kernel take on average 2ms and on the jetson it takes 8ms.

Hello, xav123:
Would you please share your test pattern? It’s based on openCV or other libs?

br
ChenJian

What do you mean by share pattern?

It is based on personnal implementaion of sobel that I shared on the git link that I put on the previous post.

Hello, xav123:
You mean the code @ https://github.com/xav12358/SobelK1?
Sorry that I missed that.

Let me try it in my side, and give you update once it’s done.

br
Chenjian

Indeed there are the two codes.

One on the Shield tablet and the other on the Jetson

Hello, xav123:
I tried your code (I have no Shield tablet in hand, the following results come from Jetson TK1.).

Before CPU/GPU/EMC clock boosted, ./CCL_gpu results (the first column statistics)

476 times, sum 5617.38886 average 11.80123710084033613445; max 20.58941; min 3.90158

After CPU/GPU/EMC clock boosted (4 CPU–2.3205GHz GPU–852MHz EMC–924MHz, tegrastats can show current clock)

./CCL_gpu results

476 times, sum 1022.11251 average 2.14729518907563025210; max 2.89817; min 2.06983

br
Chenjian

Thank you for the reply.

How can I boost the CPU/GPU/EMC clock ?

http://elinux.org/Jetson/Performance shows CPU/GPU clock control.

For EMC clock, you can refer to http://elinux.org/Jetson/TX1_Controlling_Performance, though it’s for TX1, same sysfs node should exist in TK1 as well.

br
Chenjian

I apply the modifications and it seems to work properly.

Thank you.

I also have a trouble with the 2d texture interpolation.

I made an other post for that if you could take a look at:

https://devtalk.nvidia.com/default/topic/959143/image-interpolation/#4957012