search

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

Issue 51 OpenCl kernel compilation errors for android #52

Closed DVEfremov closed 7 years ago

DVEfremov commented 7 years ago

In fact caffe uses DType as float but looks like double type support not correctly implemented: there are kerenel compilation errors on startup atleast for Mali GPU even if DOUBLE_SUPPORT_AVAILABLE not defined, so it looks like usages of DOUBLE_SUPPORT_AVAILABLE not for all required parts of code.

naibaf7 commented 7 years ago

Uh I'm not sure this is fine, as Dtype should be used preferentially. We can't modify the code like that only because the Mali GPU has a crippled implementation, I think.

psyhtest commented 7 years ago

The problem is elsewhere. I will hunt for a better solution...

DVEfremov commented 7 years ago

Yes sure I'm totally agree. We will find better aproach for that (tomorrow I'll try to check if DOUBLE_SUPPORT_AVAILABLE is ok). Just may be you have heard from some one else something related to this problem.

naibaf7 commented 7 years ago

@DVEfremov Indeed, we had such problems before. It's mostly down to horrible OpenCL implementations. Some functions might just not be defined for double.

psyhtest commented 7 years ago

@naibaf7 I'm pretty sure this particular OpenCL implementation doesn't have such gaping holes :). But let's see what solution @DVEfremov has found...

DVEfremov commented 7 years ago

I've provided new PR https://github.com/naibaf7/caffe/pull/57 with new solution @naibaf7 @psyhtest Please take a look

I'm closing this one.