etl-gpu-blas (egblas)

Mini BLAS-like library for GPU (complementary to CUBLAS).

The goal of this library is principally to be used as a complement to CUBLAS in the ETL library. The goal is to add functions that are not present in CUBLAS and make them available in the same format.

Disclaimer: All the functions are mostly expecting row-major input All functions with more than 2D are always row-major.

Features

So far, the library supports the following features:

• Vector sum (egblas_Xsum)
• Vector scalar addition (egblas_scalar_Xadd)
• Vector scalar division (egblas_scalar_Xdiv)
• Vector element-wise sqrt (egblas_Xsqrt)
• Vector element-wise log (egblas_Xlog)
• Vector element-wise exp (egblas_Xexp)
• y = (alpha * x) * y (egblas_Xaxmy)
• y = (alpha * x) / y (egblas_Xaxdy)

All functions are supporting single-precision floating points (s) and double precision floating points (d). When possible, the functions are also supporting single precision complex floating points (c) and double precision complex floating points (z).

Synchronization

By default, most of the kernels executed by this library are not synchronized. In the future, no kernel will be synchronized. If you want to synchronize after the function call, you can use cudaDeviceSynchronize() after the egblas function call. If you want all egblas functions to be synchronized, you can define EGBLAS_SYNCHRONIZE:

EXTRA_CXX_FLAGS=-DEGBLAS_SYNCHRONIZE make

In that case, every egblas function will be terminated by a cudaDeviceSynchronize() call. This can have a big performance impact, especially if working on small collections of data, since the kernel launch has a high overhead.