Tensor Library

---

C/C++ template library for tensor operations. All tensors are based on C/C++ std::vectors.

All the signatures can be checked by looking into the header files. The libraries does not automatically include the required STL dependencies. All the header files can be individually included, or jointly included by:

#include "tensor_lib.h"

Tensor Compare

tensor_lib::compare2Darray<type>(arrayA, arrayB, Ysize, Xsize) // Compares 2 C-style arrays, returns true if equals
tensor_lib::compareTensors<type>(tensorA, tensorB)             // Compares 2 tensors, accepts 1-to-4 dimensional vectors

Tensor Display

tensor_lib::print<type>(tensor) // Prints the given tensor, accepts 1-to-4 dimensional vector

Tensor Generate

tensor_lib::maxBitVal(bitwidth)                           // Returns the maximum value of a given bitwidth
tensor_lib::minNegBitVal(bitwidth)                        // Returns the minimum 2's complement value of a given bitwidth
tensor_lib::randTensor<type, bitwidth>(tensor)            // Randomizes the tensor's data with Integral value of a given bitwidth
tensor_lib::randTensor<type>(tensor, lowBound, highBound) // Randomizes the tensor's data with Float values in the range lowBound-highBound

Tensor Debug

tensor_lib::debug::linear<type>(inVec, wMat, oVec)      // Performs 2D lienar layer between 2D inputs
tensor_lib::debug::convolution<type>(inTen, wTen, oMap) // Performs 2D convolution layer between 4D inputs
tensor_lib::debug::bias_add<type>(inF, Bias, oF)        // Performs 1D vector bias addition on 2D or 4D tensors
tensor_lib::debug::ew_add<type>(A, B, Out)              // Performs Element-Wise addition over 2D or 4D tensors
tensor_lib::debug::relu<type>(inF, oF)                  // Performs ReLU over 2D or 4D input tensors
tensor_lib::debug::max_pool<type>(inF, oF, KernelSize)  // Performs a MaxPool of a squared kernel over 4D input tensors

Tensor Quant

See QUANT.md for an explanation of the basic quantization mechanics

tensor_lib::quantization::MaxMinObserver<type>(tensor, min, max)                      // Implements a MinMax observer on 1D-to-4D tensors
tensor_lib::quantization::AffineCalibration<type>(min, max, bitwidth, scale, zero)    // Implements an Affine Calibration
tensor_lib::quantization::SymmetricCalibration<type>(min, max, bitwidth, scale, zero) // Implements an Symmetric Calibration
tensor_lib::quantization::Quantize<fType, iType>(inT, oT, scale, zero)                // Quantiza the 1D-to-4D input tensor
tensor_lib::quantization::Dequantize<iType, fType>(inT, oT, scale, zero)              // Dequantize the 1D-to-4D input tensor