OpenCL error -13 on Ubuntu 14.04 and beignet

[OliverUrbann]

Used current clBLAS as the provided .so leads to many linker errors. darknet_cl compiles now but executing

build/darknet_cl detect cfg/yolo.cfg yolo.weights data/dog.jpg

leads to some warnings and -13:

31 detection
mask_scale: Using default '1.000000'
Beignet: "unable to find good values for local_work_size[i], please provide\n" " local_work_size[] explicitly, you can find good values with\n" " trial-and-error method."
Loading weights from yolo.weights...Done!
im2col_kernels.cl build log:
stringInput.cl:36:18: warning: '/*' within block comment

activation_kernels.cl build log:
stringInput.cl:21:12: warning: double precision constant requires cl_khr_fp64, casting to single precision

opencl execution error, code -13 -13
terminate called after throwing an instance of 'std::runtime_error'
what(): OpenCL error, code: -13
[1] 9856 abort (core dumped) build/darknet_cl detect cfg/yolo.cfg yolo.weights data/dog.jpg

[ganyc717]

Hi @OliverUrbann
Error code -13 means CL_MISALIGNED_SUB_BUFFER_OFFSET, means when you create a sub buffer from a GPU buffer, there is limitations on the offset alignment. This is hardware limitations and depending on GPU vendor. Currently I have no good way to resolve this issue as it will increase the complexity for the data layout on memory. Anyway, I only met such problem when using OpenCL running on CPU.
Best Regards!

[OliverUrbann]

I tried it on a Atom J1900 with z3700 GPU in an embedded system. We cannot use other GPUs so the OpenCL port would be a great way to execute YOLO on that system. Anyway, thx for the explanation.

[mhahn0106]

@ganyc717 I managed to compile and run darkness_cl on AMD Radeon GPU by adding some memory alignment codes but getting wrong detection result compared with CPU version. I hope you to explain design of CLArray. Can you leave some notes for it? To my understanding, it seems like making slices of given large continuous memory block. If we can focus on continuity, how about using SVM(Shared Virtual Memory) from OpenC L2.0?

[ganyc717]

Hi @mhahn0106
CLArray is a ugly design as I have no idea to random access to the cl_mem buffer in CL 1.2. I am not quite familiar with OpenCL 2.0 new APIs. Thank you for your idea, I will check the spec and follow your idea later.
Best Regards!

[ganyc717]

Hi @OliverUrbann @mhahn0106
I have create a new branch enable-shared-virtual-memory, this branch using shared virtual memory instead of cl_mem. I tested it on Intel(R) Core(TM) i5-4570 CPU with opencl, and won't met CL_MISALIGNED_SUB_BUFFER_OFFSET again.
I kept the CLArray design as the clBLAS APIs need cl_mem instead of svm pointers.
Best Regards!

[mhahn0106]

Awesome. Looks good. Will test on my box and let you know.

[OliverUrbann]

That's great, thx a lot for your efforts.

However:

-- Looking for CL_VERSION_2_1 - not found
-- Looking for CL_VERSION_2_0
-- Looking for CL_VERSION_2_0 - found
-- Found OpenCL: /usr/lib/x86_64-linux-gnu/libOpenCL.so (found version "2.0")

So I have OpenCL 2.0. And get:

gemm.cpp:(.text+0x77f): Nicht definierter Verweis auf clEnqueueSVMMap' gemm.cpp:(.text+0x7c4): Nicht definierter Verweis auf clEnqueueSVMMap'
gemm.cpp:(.text+0x808): Nicht definierter Verweis auf clEnqueueSVMMap' gemm.cpp:(.text+0x9d9): Nicht definierter Verweis auf clEnqueueSVMUnmap'
gemm.cpp:(.text+0x9fd): Nicht definierter Verweis auf clEnqueueSVMUnmap' gemm.cpp:(.text+0xa21): Nicht definierter Verweis auf clEnqueueSVMUnmap'

It's german for undefined reference to. I'm a bit confused. According to the beignet doc OpenCL 2.0 is only supported on Skylake and up, so that would explain this, abut not why OpenCL 2.0 is found.

[soulslicer]

@ganyc717

"CL_MISALIGNED_SUB_BUFFER_OFFSET, means when you create a sub buffer from a GPU buffer, there is limitations on the offset alignment."

Can you explain what this means to me in more detail. What do you mean by limitations of offset alignment. Can you provide a simple example?

[ganyc717]

@soulslicer
You have a cl_mem buffer, and wanna to create a sub buffer on it, but it does not mean an arbitrary offset is acceptable. You can refer to https://www.khronos.org/registry/OpenCL/sdk/1.1/docs/man/xhtml/clCreateSubBuffer.html

[chenxian9999]

@ganyc717
Thx a lot for your work.
I run branch Master on Win10 with a AMD GPU, and i recieved a error -13.
then，i run on branch: test-enable-shared-virtual-memory, i recieved error -14 in void fill_gpu(int N, float ALPHA, CLArray X, int INCX)
So what should I do next?

[ganyc717]

@chenxian9999
I enable the shared virtual memory on different platforms, some works but some failed with error -14. I will find why get such error.

[chenxian9999]

Thx a lot.
Looking forward to your improvement.

[aaronsta1]

i too am getting error -13 has there been a fix yet?
i did get this one to compile, but its an old build
https://github.com/myestro/darknet

`im2col_kernels.cl build log:
/tmp/OCL2394T5.cl:36:18: warning: '/*' within block comment
//*data_col_ptr = data_im_ptr[ii * width + jj];
^
1 warning generated.

opencl execution error, code -13 -13
terminate called after throwing an instance of 'std::runtime_error'
what(): OpenCL error, code: -13
`