ArgMax operation (GPU Implementation)

[vanmaxim]

[vanmaxim]

@nolanliou @llhe Hi. Thanks for your amazing work! I need some help. I don't get how to make an output tensor with int type? Also how to make an output tensor with three dimensions (like for CPU implementation)?

[yejw5]

@vanmaxim Thanks for your contributions!

• Could you show that why you need argmax operation on GPU? Is there some shortage in the CPU implementation?
• OpenCL is not good at controlflow statements, such as if-else, the same in CUDA. So the running may cost more time than CPU.
• We haven't support 3-D memory layout for OpenCL Image struct, refer to: https://mace.readthedocs.io/en/latest/development/memory_layout.html. It seems to be hard for mapping dimensions in tensors to Images.

[vanmaxim]

@yejw5 How can I make output tensor with int type? PS. I'll prepare performance benchmarks.

[yejw5]

@vanmaxim Did you mean model's output in MaceTensor::data()？ Currently, only float data type is supported. But you can reinterpret_cast the data pointer as int type.