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. 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

./bench/bench-blas OP

where OP is axpy, gemv or gemm. It detects clBLAS or cuBLAS and compares it against ISAAC for DeepBench, Covariance and LAPACK (packed rank1 updates).

Here's what you get for a Pascal Titan X (TFLOPS):

BENCH	M	N	K	AT	BT	ISAAC	cuBLAS
Deep	1760	16	1760	N	N	1.31	1.65
-	1760	32	1760	N	N	2.29	1.71
-	1760	64	1760	N	N	3.35	2.34
-	1760	128	1760	N	N	5.04	4.46
-	1760	7000	1760	N	N	8.16	10.10
-	2048	16	2048	N	N	1.60	1.30
-	2048	32	2048	N	N	2.80	1.93
-	2048	64	2048	N	N	3.69	2.20
-	2048	128	2048	N	N	4.84	5.07
-	2048	7000	2048	N	N	8.30	10.39
-	2560	16	2560	N	N	1.87	1.27
-	2560	32	2560	N	N	3.45	2.25
-	2560	64	2560	N	N	4.82	2.71
-	2560	128	2560	N	N	5.08	6.25
-	2560	7000	2560	N	N	8.32	10.11
-	1760	16	1760	T	N	0.76	0.62
-	1760	32	1760	T	N	1.45	1.39
-	1760	64	1760	T	N	2.38	2.36
-	1760	128	1760	T	N	3.20	2.46
-	1760	7000	1760	T	N	6.89	8.94
-	2048	16	2048	T	N	0.99	0.88
-	2048	32	2048	T	N	1.97	2.18
-	2048	64	2048	T	N	2.77	2.33
-	2048	128	2048	T	N	3.08	4.50
-	2048	7000	2048	T	N	7.17	6.89
-	2560	16	2560	T	N	0.86	0.71
-	2560	32	2560	T	N	1.69	1.73
-	2560	64	2560	T	N	2.57	2.49
-	2560	128	2560	T	N	2.66	2.93
-	2560	7000	2560	T	N	7.03	4.62
-	1760	7133	1760	N	T	8.65	10.02
-	4096	7133	4096	N	T	8.98	9.73
Cov	32	60000	32	N	T	1.18	0.74
-	256	60000	256	N	T	6.78	3.21
Lapack	4096	4096	32	N	T	4.17	2.58
-	3456	3456	32	N	T	3.90	2.50
-	896	896	32	N	T	1.25	1.34

Here's what you get for an AMD Fury (TFLOPS):

BENCH	M	N	K	AT	BT	ISAAC	clBLAS
Deep	1760	16	1760	N	N	0.21	0.13
-	1760	32	1760	N	N	0.43	0.27
-	1760	64	1760	N	N	0.84	0.53
-	1760	128	1760	N	N	1.69	0.99
-	1760	7000	1760	N	N	3.59	2.71
-	2048	16	2048	N	N	0.24	0.16
-	2048	32	2048	N	N	0.52	0.31
-	2048	64	2048	N	N	0.88	0.57
-	2048	128	2048	N	N	1.43	0.87
-	2048	7000	2048	N	N	3.47	2.49
-	2560	16	2560	N	N	0.47	0.19
-	2560	32	2560	N	N	0.94	0.38
-	2560	64	2560	N	N	1.59	0.73
-	2560	128	2560	N	N	1.94	1.22
-	2560	7000	2560	N	N	3.65	2.65
-	1760	16	1760	T	N	0.21	0.52
-	1760	32	1760	T	N	0.41	0.89
-	1760	64	1760	T	N	0.81	1.10
-	1760	128	1760	T	N	1.41	1.28
-	1760	7000	1760	T	N	2.87	1.73
-	2048	16	2048	T	N	0.23	0.37
-	2048	32	2048	T	N	0.42	0.40
-	2048	64	2048	T	N	0.54	0.72
-	2048	128	2048	T	N	0.68	0.95
-	2048	7000	2048	T	N	1.96	1.29
-	2560	16	2560	T	N	0.41	0.67
-	2560	32	2560	T	N	0.83	0.81
-	2560	64	2560	T	N	0.92	1.10
-	2560	128	2560	T	N	1.07	1.14
-	2560	7000	2560	T	N	2.36	1.49
-	1760	7133	1760	N	T	3.47	2.70
-	4096	7133	4096	N	T	3.54	2.61
Cov	32	60000	32	N	T	0.42	0.00
-	256	60000	256	N	T	1.71	0.19
Lapack	4096	4096	32	N	T	1.38	1.31
-	3456	3456	32	N	T	1.54	1.24
-	896	896	32	N	T	0.37	0.34

BLAS routines supported

Currently supported functions are:

BLAS1	BLAS2	BLAS3
xAXPY	xGEMV	xGEMM
xCOPY	xGER
xSCAL
xDOT
xASUM

Contributing

Non-tuned GPUs are:

• Intel Skylake iGPU
• AMD GCN > 1.2
• Intel Xeon Phi

I'm planning on adding double precision support very soon.

I'm no longer very active on this project, although I'll fix bugs. I've been working on a more ambitious project lately.