Is Flex delegate compatible with CUDA?

[Iosif Hamlatzis]

I have created a model with SELECT OPS but if I have CUDA enabled it crashes both under Python and C++, but works fine if CUDA isn't enabled

When CUDA is enabled (and only then) I get an error message 

"op 'ReadVariableOp' output%d tensor#%d shape mismatch for [144] != []" from file "./tensorflow/lite/delegates/flex/kernel.cc" at line 590 and then it goes to file "./tensorflow/lite/delegates/flex/delegate.cc" in method "CopyFromBufferHandle" in line 160

and crashes "memcpy(output->data.raw, t_data.data(), t_data.size());"

Any idea?

[Terry (Woncheol) Heo]

Since the main target of TFLite is mobile and IoT devices, CUDA is not our major target environment.

So the using CUDA for SELECT OPS is not verified.

But if you're interested in GPU acceleration of mobile, you could use OpenCL for TFLite kernels.

https://www.tensorflow.org/lite/guide/build_cmake#opencl_gpu_delegate

Thanks.

Terry