wujunkai166/isaac

Input-aware cuBLAS/clBLAS implementation for better performance

ISAAC

This is the developer repository for ISAAC, a library that uses machine learning to find input-aware kernels for element-wise operations, 1D/2D reductions and GEMM. It works with both cuBLAS and clBLAS. It's super easy to compile (no dependency!), to install (just link against libisaac.so instead of clBLAS or cuBLAS!), almost always outperforms (tuned) clBLAS and often outperforms cuBLAS. And when it predict that it doesn't, it fallbacks on vendor libraries automatically. Try it!

License

ISAAC is distributed under the GNU LGPL v2.1 License.

Installation

ISAAC is dependency-free, and will load either OpenCL and/or CUDA 7.0+ dynamically depending on which GPUs are detected at runtime.

You only need CMake 2.8.7+ and a C++11 compliant compiler:

git clone https://github.com/ptillet/isaac.git
mkdir -p isaac/build && cd isaac/build
cmake ../ && make -j4

Link against libisaac.so instead of libcublas.so or libclblas.so, and you're good to go!

The C++ and Python API does some kernel fusion, but is not entirely stable. It works well to compose element-wise operations, though.

Benchmark

Usage : blas-bench [--op {axpy, dot, gemv, gemm}] [--dtype {float32, float64}] [--device DEVICE_IDX] [--help]
--op: operation to benchmark (default = gemm)
--dtype: data-type to benchmark (default = float32)
--device: index of isaac device in [0, ..., ndevices - 1] (default = 0)
--help: display this message

It detects clBLAS or cuBLAS and compares it against ISAAC for e.g., DeepBench, Covariance, LAPACK (packed rank1 updates), etc.

Below is the TFLOPS you get for GEMM on a Pascal Titan X (cuBLAS 8.0). Numbers in bold represent speed-ups greater than 5%.

For AMD Fury (clBLAS-2.10-Fiji):

Same trend on Intel Broadwell iGPU

BLAS routines supported

Currently supported functions are:

BLAS1	BLAS2	BLAS3
xAXPY	xGEMV	xGEMM
xCOPY	xGER
xSCAL
xDOT
xASUM

For x in {S, D}

Contributing

You can contribute to further tuning isaac if you have one of the following architecture:

• NVidia: SM 2.x ; SM 3.5 ; SM 5.0
• Intel: Skylake iGPU

If you have one of the following architectures you can contribute by running:

git clone https://github.com/ptillet/isaac.git
cd isaac/python ;
python setup.py build;
cd ../tune
PYTHONPATH=../python/build/lib.linux-x86_64-2.7/ python main.py --float32 --elementwise_1d --elementwise_2d --reduce_1d --reduce_2d_rows --reduce_2d_cols --gemm_nn --gemm_nt --gemm_tn --gemm_tt

This will output a .json file that you can submit for integration. Add --float64 if your GPU has double-precision capabilities.

Bug reports are more than welcome!