accelerating machine learning using blis - home -...
TRANSCRIPT
Accelerating Machine Learning using BLIS
Santanu Thangaraj, Kiran Varaganti, Kiran Puttur, Pradeep Rao
Advanced Micro Devices, Inc
Introduction: Taking advantage of low latency and hierarchical memory architecture of x86 is
critical to boost the performance of computational intensive applications such as deep
learning algorithms in AMD platforms. Machine Learning (ML) algorithms are primarily
built on top of basic linear algebra subprograms (BLAS). Hence performance of these
linear algebra routines directly impact the performance of ML algorithms. In our
experiments we use Caffe [4], a deep learning framework implementation and compare its
performance by linking against BLAS libraries such as BLIS [7] and OpenBLAS [9].
Existing BLIS library performs poorly when benchmarked with Caffe’s handwritten
digits recognition (MNIST challenge [1], [10]) deep layer model. We addressed this
shortcoming in BLIS library and optimize the library to perform better for machine learning
frameworks. We refer to optimized BLIS library as AMD optimized BLIS library
BLAS specifications:
BLAS is a specification that prescribes a set of low-level routines for performing
common linear algebra operations such as vector addition, scalar multiplication, dot
products, linear combinations, and matrix multiplication. They are the de facto standard
low-level routines for linear algebra libraries; the routines have bindings for both C and
FORTRAN. Although the BLAS specification is general, BLAS implementations are often
optimized for speed on a particular machine, so using them can bring substantial
performance benefits. BLAS implementations will take advantage of floating point
hardware such as vector registers and SIMD instructions. Examples of BLAS libraries
include: OpenBLAS, University of Texas Austin’s BLAS-like Library Instantiation
Software Framework (BLIS) and Intel Math Kernel Library (MKL).
Basic linear algebraic operations exposed by BLAS libraries forms the crucial
component of Machine Learning algorithms. Many machine learning frameworks
including Caffe [10], depend on BLAS libraries to provide the required linear algebra
functionality and can link to any of the standard BLAS libraries.
AMD adopted BLIS as its new BLAS library. AMD will provide optimized BLIS
library for their microprocessors based on the new x86 architecture codenamed “Zen”.
BLIS framework was designed to isolate essential kernels of computation that, when
optimized, immediately enable optimized implementations of most of its commonly used
and computationally intensive operations. BLIS is written in ISO C99 and available under
a new/modified/3-clause BSD license.
Deep Learning with Caffe Convolutional Neural Networks (CNNs) are successful class of DNNs. CNNs are
computed using dense kernels that differ from traditional dense linear algebra routines.
Accordingly, modern deep learning frameworks such as Caffe provides suites of custom
kernels that implement basic operations such as tensor convolutions, activation functions
and pooling. These routines represent the bulk of the computations when training a CNN,
and thus account for the majority of its execution time. The deep learning community has
been successful in finding optimized implementations of these kernels, but as the
underlying architectures evolve, these kernels must be re-optimized, which is a significant
investment. Optimizing these kernels requires a deep understanding of the underlying
processor architecture, with careful scheduling of data movement, on-chip memory
placement, register blocking, and other optimizations in order to get acceptable
performance.
Role of BLAS in DNN: The most important computational primitive in CNNs is a special form of batched
convolution called spatial convolution [1] [5].
There are two inputs to the convolution: 𝐷 ∈ ℝ𝑁𝐶𝐻𝑊 , which forms the input data,
and 𝐹 ∈ ℝ𝐾𝐶𝑅𝑆, which forms the convolutional filters. The input data ranges over N images
in a mini batch, C input feature maps, H rows per image, and W columns per image. The
filters range over K output feature maps, C input feature maps, R rows per filter, and S
columns per filter. Computing this convolution involves a seven-way nested loop, with
four independent loops and three accumulation loops [5]. There are many ways of
implementing this computation. The Caffe MNIST benchmark training algorithm
implements by lowering the convolutions onto a matrix multiplication (GEMM). The
GEMM gets invoked for small matrix sizes. Therefore the performance of small matrix
GEMM directly impacts the performance of the training algorithm. The optimized GEMM
implementations are provided by BLAS libraries.
Small matrix GEMM optimization:
The BLIS library has six loops [12] around the GEMM computation, with the outer
loop parameters dependent on L3 cache size while the inner loops dependent on L1/L2
cache sizes. The packing of data, required by inner loops, is done to avoid TLB misses.
This approach gives better performance for really large matrices which does not fit entirely
in the cache system but introduces unnecessary overhead for small matrix computations.
We have optimized GEMM specifically for small matrix cases and observed significant
performance improvements (refer figure 1).
For our benchmarks we have used the Caffe version 1.0.0.rc3, OpenBLAS 0.2.20,
BLIS 0.2.1 public open-source repository and AMD optimized BLIS version x.y (TBD).
The experiments were run on Ubuntu 15.04 operating system.
Figure 1. Results of single thread SGEMM performance with BLIS public, OpenBLAS and BLIS optimized.
Machine: AMD Naples, 64 cores, 256 GB RAM @ 3.2 GHz.
Caffe MNIST performance improvement:
The optimization of GEMM the performance of the Caffe MNIST has improved, as
shown in Figure 2. The performance of the forward pass shows significant improvement
(Lower is better) and Caffe performance as a whole has improved by 17%.
0
10
20
30
40
50
60
5 20 35 50 65 80 95 110 125 140 155 170 185 200 215 230 245 260 275 290
GFL
OP
S
Matrix Size
Small Matrix - SGEMM
BLIS public Zen Optimized BLIS OpenBLAS
Average of 47% improvement
Figure 2. Results of single thread Caffe with OpenBLAS, BLIS Public and BLIS optimized. Machine: AMD
Zen Naples, 64 cores, 256 GB RAM @ 3.2 GHz.
*lesser the time, better is the performance
Looking at the backward pass performance numbers there is a room for further
improvement. The reason is the significant amount of GEMM calls made during the
backward pass requires transpose of the input matrices. This support is not supported yet
by the small matrix code.
Conclusion: Machine learning, Deep Neural Networks are significantly gaining traction across
the Industries for its application in automating every day chores and bringing AI into
everyday life. Most of the Machine learning frameworks links with BLAS libraries during
compilation. BLAS forms the de facto standard low-level routines for linear algebra
libraries. It is the layer upon which lot of other high level Dense Linear Applications (DLA)
are based. Having highly optimized BLAS library is essential for accelerating the ML
frameworks.
By optimizing the level 1 and level 2 subroutines and small matrix GEMM we are
able to achieve significant boost in performance of Caffe run with MNIST. We could
observe performance benefit is not limited to only Caffe, but also for the LAPACK [8]
routines such as LU, QR and Cholesky.
21.3 19.3 17.8
25.523.5 23.3
0
10
20
30
40
50
BLIS Public OpenBLAS Zen OptimizedBLIS
Tim
e in
ms
BLAS library
Caffe MNIST
forward backward
17% improvement
Reference:
[1] Yann LeCun, Leon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based
learning ´ applied to document recognition. Proceedings of the IEEE, 86(11):2278–
2324, 1998.
[2] Ronan Collobert, Koray Kavukcuoglu, and Clement Farabet. Torch7: A matlab-like
environ- ´ ment for machine learning. In BigLearn, NIPS Workshop, 2011.
https://github.com/torch/torch7.
[3] James Bergstra, Olivier Breuleux, Fred´ eric Bastien, Pascal Lamblin, Razvan
Pascanu, Guil- ´ laume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua
Bengio. Theano: a cpu and gpu math expression compiler. In SciPy, volume 4, page 3,
2010. https://github.com/Theano/Theano.
[4] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross
Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture
for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014.
https://github.com/BVLC/caffe.
[5] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Cohen, John Tran,
NVIDIA cuDNN: Efficient Primitives for Deep Learning.
https://arxiv.org/pdf/1410.0759.pdf
[6] Anatomy of High-performance Matrix multiplication Kazushige Goto, the University
of Texas at Austin. Robert A. Van De Geijn, the University of Texas at Austin
https://www.cs.utexas.edu/users/pingali/CS378/2008sp/papers/gotoPaper.pdf.
[7] BLAS-like Library Instantiation Software Framework. https://github.com/flame/blis.
[8] LAPACK - Linear Algebra PACKage : http://www.netlib.org/lapack/.
[9] OpenBLAS: An optimized BLAS library. http://www.openblas.ne
[10] Caffe:http://caffe.berkeleyvision.org/gathered/examples/mnist.html.
[11] INTEL MKL: https://software.intel.com/en-us/intel-mkl.
[12] BLIS multithreading: https://github.com/flame/blis/wiki/Multithreading.
DISCLAIMER The information contained herein is for informational purposes only, and is subject to change without notice. While every precaution has been taken in the preparation of this document, it may contain technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties with respect to the accuracy or completeness of the contents of this document, and assumes no liability of any kind, including the implied warranties of noninfringement, merchantability or fitness for particular purposes, with respect to the operation or use of AMD hardware, software or other products described herein. No license, including implied or arising by estoppel, to any intellectual property rights is granted by this document. Terms and limitations applicable to the purchase or use of AMD’s products are as set forth in a signed agreement between the parties or in AMD's Standard Terms and Conditions of Sale. AMD, the AMD Arrow logo, and combinations thereof are trademarks of Advanced Micro Devices, Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies. © 2017 Advanced Micro Devices, Inc. All rights reserved.