caffeOnACL requires OpenCL device support?

[WhiskeyBravo]

Issue summary

Cannot link against caffeOnACL libraries when ACL is compiled without opencl. My target device, RPi 3, has a CortexA53 and no OpenCL runtime, but would benefit from NEON acceleration. Thoughts?

Steps to reproduce

Compile ACL
scons -j3 debug=0 asserts=0 arch=arm64-v8a os=linux Werror=1 build=cross_compile neon=1 opencl=0 embed_kernels=0 openmp=0 cppthreads=1
Compile caffeOnACL

-DBUILD_docs:BOOL=OFF
-DBUILD_SHARED_LIBS:BOOL=OFF
-DCPU_ONLY:BOOL=ON
-DUSE_CUDNN:BOOL=OFF
-DBUILD_matlab:BOOL=OFF
-DBUILD_python:BOOL=OFF
-DBUILD_python_layer:BOOL=OFF
-DUSE_LMDB:BOOL=OFF
-DUSE_LEVELDB:BOOL=OFF
-DUSE_NCCL:BOOL=OFF
-DUSE_OPENCV:BOOL=OFF
-DBLAS:STRING=Open
-DUSE_ACL:BOOL=ON
.....other flags truncated

Your system configuration

Operating system: Host OS Linux, target OS Linux
Compiler: aarch64-linux-gnu-g++-6

[AnthonyBarbier]

I don't know much about the Caffe integration, but you can compile the Compute Library with OpenCL enabled: it uses dlopen to resolve OpenCL symbols at runtime when the first OpenCL entry point is called so if your system doesn't have OpenCL it should still work.

Hope this helps

[WhiskeyBravo]

@AnthonyARM thanks for pointing me in this direction, I got it running but performance is approx 1.7x slower than compiling caffe and all it's dependencies from source with optimizations turned on and targetting mcpu=cortex-a53
I am running essentially ResNet. Thoughts?

[mihailescu2m]

I got similar results.
Running caffe-on-acl (ogl enabled) was almost 2x slower than running caffe
Odroid XU4/Mali T628 r12 driver

[kuang93]

@mihailescu2m I want to running caffe on Odroid XU4 with GPU mode， but I can only find few tutorials about this. How do you achieve this? thanks :)