naibaf7 / caffe

Caffe: a fast open framework for deep learning. With OpenCL and CUDA support.
http://caffe.berkeleyvision.org/
Other
85 stars 20 forks source link

Mali GPU does not support host unified memory in fact #53

Closed DVEfremov closed 7 years ago

DVEfremov commented 7 years ago

Step to reproduce: 1) Apply PR #50 and PR #51 based on opencl branch to get android version of caffe binary with required libs

2) run caffe time comand and got error

and strange artifacts on the screen while processing caffe command atleast for Galaxy Note 4 with Mali GPU:

I0202 16:24:37.476721  5890 common.cpp:405] Device id:                     0
I0202 16:24:37.476784  5890 common.cpp:407] Device backend:                OpenCL
I0202 16:24:37.476850  5890 common.cpp:409] Backend details:               ARM: OpenCL 1.1 
I0202 16:24:37.476951  5890 common.cpp:411] Device vendor:                 ARM
I0202 16:24:37.477020  5890 common.cpp:413] Name:                          Mali-T760
I0202 16:24:37.477088  5890 common.cpp:415] Total global memory:           2147483648

So as solution provide compilation param DISABLE_DEVICE_HOST_UNIFIED_MEMORY to force disabling support host unified memory

DVEfremov commented 7 years ago

https://github.com/naibaf7/caffe/pull/54 merged to master so closing.

bhack commented 7 years ago

But is Mali requiring CL_MEM_ALLOC_HOST_PTR ?

bhack commented 7 years ago

@naibaf7 https://github.com/naibaf7/libdnn/issues/17