tblas's Introduction

tBLAS

A header only multithreaded and cache-optimized linear algebra library.

TODO and future

Optimized kernel for transpose: The current implementation supports fast transpose using tiling and multithreading for large matrices. However the size of the kernel needs to be tuned specifically for transpose. The lack of this leads to longer compute times than the naive version for transpose of large matrices.
Tuning params: Presently the parameters for block tiling the matrices are defined in constants.hpp. These values are not tuned and should be interpreted as sensible defaults. In an ideal scenario these would be tuned for a given CPU with known cache sizes. The current implementation can be improved by allowing configuration by passing preprocessor definitions so the user of the library can input the cache sizes and the block sizes can then be derived from these by using constexpr based on some heuristics.
More tests: The current tests showcase the correctness of the solution fairly well. There is still room for more coverage.

Installation

From project root run

$ cmake .. -DCMAKE_INSTALL_PREFIX:PATH=/your/installation/path
$ cmake --build . --config Release --target install -- -j $(nproc)

Running tests

From build directory run

$ cmake ..
$ make tests
$ ctest

Running benchmarks

From build directory run.

$ cmake ..
$ make bench
$ ./tests/bench  --benchmark-samples=5

Usage

Example CMake snippet

project("tBLAS-example")

find_package(tBLAS CONFIG REQUIRED)

add_executable(tBLAS-example src/main.cpp)
target_link_libraries(tBLAS-example tBLAS::tBLAS)

Recommend Projects