jaiaid / tensorswift

A very basic tensor engine done as part of course work in cisc830

TensorSwift: High-Performance Tensor Arithmetic

A very basic tensor engine done as part of course work in cisc830

Prerequisite

sudo apt-get install python3-dev

Check if pybind exists

python3 -m pybind11 --includes

If not exists

pip3 install --no-cache-dir pybind11

Build

To build

with openmp support (for cpu acceleration)

make OMP=1

without openmp support

make

To control number of thread employed for parallelization, do following modification in ts.h

#define SYS_PARAM_CPUCOUNT 8

change the number according to your system's core count

Output will be an executable file tensorswift.cpython-38-x86_64-linux-gnu.so (depends on python version)

To clean outputs

make clean

Example usage

From the same folder where tensorswift.cpython-38-x86_64-linux-gnu.so exists

Check if tensorswift found,

python3 -c "import tensorswift"

Currently two class exists in the module,

1. Data (a simple int wrapping class with basic four operation support)
2. SwiftTensor, ongoing work, buffer contains garbage, size is in uint64_t and means number of element, arbitrary dimension should work, only floating number hold

Example import

a = tensorswift.Data(1)
b = tensorswift.Data(2)

print(a)
print(b)

c = a+b

# empty tensor
a = tensorswift.SwiftTensor()
# tensor with shape [1,2]
b = tensorswift.SwiftTensor([1,2])

example code can be found in example.py

tensorswift's Issues

Number of bits used for the buffer size

The buffer size may be larger than 32 bits. Thus, using int may result in a truncated buffer. We can change it to uint64_t.