Each Tensor instance is a multi-dimensional array allocated on a specific Device instance. Tensor instances store variables and provide linear algebra operations over different types of hardware devices without user awareness. Note that users need to make sure the tensor operands are allocated on the same device except copy functions.
SINGA has three different sets of implmentations of Tensor functions, one for each type of Device.
- ‘tensor_math_cpp.h’ implements operations using Cpp (with CBLAS) for CppGPU devices.
- ‘tensor_math_cuda.h’ implements operations using Cuda (with cuBLAS) for CudaGPU devices.
- ‘tensor_math_opencl.h’ implements operations using OpenCL for OpenclGPU devices.