AlexNet: libGLES_mali.so & libmali.so missing

[Napoleon_BlownApart]

I'm trying the Running AlexNet on Raspberry Pi with Compute Library
tutorial on my Raspberry Pi 3 but after successfully completing all the preparatory steps, the test fails to classify the 'go-kart' image.

On my Pi, I have Ubuntu 16.04.5 LTS installed (not Mate). Details below:

Code: Select all

user@pi:[23:21]:~$ cat /proc/cpuinfo
...
Hardware	: BCM2709
Revision	: a22082
user@pi:[23:21]:~$ lsb_release -a
No LSB modules are available.
Distributor ID:	Ubuntu
Description:	Ubuntu 16.04.5 LTS
Release:	16.04
Codename:	xenial

I compiled the ComputeLibrary with 'opencl=1' but removed '-j2' as it was misinterpreted as a make target, using:

Code: Select all

scons Werror=1 debug=0 asserts=0 neon=1 opencl=1 examples=1 build=native

Running the 'go-kart' test produced the following result, which is incorrect as it didn't classify the image:

Code: Select all

user@pi:[22:17]:~/Programming/CNN/ComputeLibrary$ ./build/examples/graph_alexnet 0 $PATH_ASSETS $PATH_ASSETS/go_kart.ppm $PATH_ASSETS/labels.txt

./build/examples/graph_alexnet

Threads : 1
Target : NEON
Data type : F32
Data layout : NHWC
Tuner enabled? : false
Tuner file :
Fast math enabled? : false

Can't load libOpenCL.so: libOpenCL.so: cannot open shared object file: No such file or directory
Can't load libGLES_mali.so: libGLES_mali.so: cannot open shared object file: No such file or directory
Can't load libmali.so: libmali.so: cannot open shared object file: No such file or directory
Couldn't find any OpenCL library.

Test passed

From the error I realised that I didn't have OpenCL installed and possibly OpenGLES either, so I installed the following

Code: Select all

sudo apt-get install ocl-icd-opencl-dev ocl-icd-dev
sudo apt-get install cmake libx11-dev xorg-dev libglu1-mesa-dev freeglut3-dev libglew1.5 libglew1.5-dev libglu1-mesa libglu1-mesa-dev libgl1-mesa-glx libgl1-mesa-dev

I now have /usr/lib/arm-linux-gnueabihf/libOpenCL.so but I don't have the other two. However, I do have /usr/lib/libGLESv2.so.

In the mean time, I'm rebuilding with 'opencl=0' to see if it works without the GPU boost, but I really want to utilise the VideoCore IV.
Update : Running the test after rebuilding produced:

Code: Select all

user@pi:[ 1:45]:~/Programming/CNN/ComputeLibrary$ ./build/examples/graph_alexnet 0 $PATH_ASSETS $PATH_ASSETS/go_kart.ppm $PATH_ASSETS/labels.txt

./build/examples/graph_alexnet

Threads : 1
Target : NEON
Data type : F32
Data layout : NHWC
Tuner enabled? : false
Tuner file :
Fast math enabled? : false

Killed

What does "Killed" mean here?

How/where can I get libGLES_mali.so and libmali.so? I've seen some advice suggesting symlinking things but my situation doesn't seem to fit what I've seen. Also, do I need to build VC4CL still?

Any help would be much appreciated.

Cheers,
Nap

[topguy]

"Killed" usually means that Linux didnt like something that your program tried to do and "killed" it.

"mali" is as far as I can remember a series of GPUs made by ARM. I'm pretty sure the Pi doesnt have one.

[Napoleon_BlownApart]

The Pi has a VideoCore IV GPU, so I'll go ahead and install VC4CL and see if that gets things going.

[Gavinmc42]

I compiled the Compute library with NEON but with no OpenCL as that only supports Mali.
I use Gentoo64, and in that Pi forum is Sakaki's how to compile it.

doe300's VC4CL might not be compatible as I don't think it is 100% OpenCL yet.

You might want to look at Tensorflow lite as that does work on Pi3's, so they say.