-
Notifications
You must be signed in to change notification settings - Fork 0
/
main.tex
882 lines (804 loc) · 45.7 KB
/
main.tex
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
\documentclass[preprint]{sigplanconf}
%\usepackage{cvpr}
\usepackage[pagebackref=true,breaklinks=true,letterpaper=true,colorlinks,bookmarks=false]{hyperref}
\usepackage{times}
\usepackage{epsfig}
\usepackage{graphicx}
\usepackage{amsmath}
\usepackage{amssymb}
\usepackage{amsfonts}
\usepackage{algorithm}
\usepackage{algorithmic}
\usepackage{subfigure}
\usepackage{pstricks}
\usepackage{tikz}
\usepackage{mydefs}
% A few things in cvpr style file but not in ACM...
\def\eg{\emph{e.g.\ }}
\def\ie{\emph{i.e.\ }}
\hyphenation{under-determined}
\begin{document}
\conferenceinfo{PPoPP '12}{Feb. 29, New Orleans.}
\copyrightyear{2012}
\copyrightdata{[to be supplied]}
\titlebanner{REVIEW COPY; PLEASE TO NOT REDISTRIBUTE} % These are ignored unless
%\preprintfooter{short description of paper} % 'preprint' option specified.
\title{Parallelization of Fast $\ell_1$-Minimization for Face Recognition}
\authorinfo{Name1}
{Affiliation1}
{Email1}
\authorinfo{Name2\and Name3}
{Affiliation2/3}
{Email2/3}
%\authorinfo{Andrew Wagner\and Yi Ma}
%{University of Illinois at Urbana-Champaign}
%{\{awagner,yima\}@illinois.edu}
%\authorinfo{Victor Shia\and Allen Yang\and Mark Murphy}
%{University of California Berkeley}
%{\{vshia,mjmurphy,yang\}@eecs.berkeley.edu}
\maketitle
\begin{abstract} Recently a family of promising face recognition algorithms
based on sparse representation and $\ell_1$-minimization ($\ell_1$-min) have been
developed. These algorithms have not yet seen commercial application,
largely due to higher computational cost compared to other traditional
algorithms. This paper studies techniques for leveraging the
massive parallelism available in GPU and CPU hardware to accelerate
$\ell_1$-min based on augmented Lagrangian method (ALM) solvers.
For very large problems, the GPU is faster due to higher memory bandwidth, while
for problems that fit in the larger CPU L3 cache, the CPU is faster.
On both architectures, the proposed implementations significantly outperform
naive library-based implementations, as well as previous systems.
The source code of the
algorithms will be made available for peer evaluation. \end{abstract}
\section{Introduction}
$\ell_1$-minimization ($\ell_1$-min) has received much attention in recent
years due to important applications in compressive sensing
\cite{BrucksteinA2007} and sparse representation \cite{WrightJ2010-PIEEE}.
The problem refers to finding the minimum $\ell_1$-norm solution to an
underdetermined linear system $\bb=A\xx$:
\begin{equation}
\min \|\xx\|_1\quad \mbox{ subj. to }\quad \bb = A\xx.
\label{eq:l1min}
\end{equation}
It is now well established that, under certain conditions
\cite{CandesE2005-IT_1,DonohoD2004}, the minimum $\ell_1$-norm solution is also
the \emph{sparsest} solution to the system \eqref{eq:l1min}.
Among its many applications, $\ell_1$-min has been recently used to reformulate
image-based face recognition as a sparse representation problem
\cite{WrightJ2009-PAMI}. If we stack the $m$ pixels of the $n_i$ training images of $K$ subject
classes into the columns of matrices $(A_1\in\Re^{m\times n_1}, \cdots, A_K\in\Re^{m\times n_K})$, combine
the matrices into a larger matrix $A = [A_1, \cdots, A_K]\in\Re^{m\times n}$, and arrange the pixels of a new
query image into a vector $\bb\in\Re^m$, \emph{sparsity-based
classification} (SBC) solves the following minimization problem:
\begin{equation}
\min_{\xx, \ee} \| \xx \|_1 + \|\ee\|_1 \quad \subj \quad \bb = A \xx + \ee.
\label{eq:l1min_denoise}
\end{equation}
If the sparsest solutions for $\x$ and $\e$ are recovered, $\ee$ provides a
means to compensate for pixels that are corrupted due to occlusion of some part of the query
image, and the dominant nonzero coefficients of $\xx$ reveal the membership of
$\bb$ based on the training image labels associated with $A$.
{\bf Several
classifiers based on $\xx$ have been proposed that all have nearly identical
performance. In this paper classification is performed using the following residual:
\begin{equation}
i^* = \argmin{i} || \bb - \ee - A_i \xx_i||_2
\label{eq:classification}
\end{equation}
}
% changed
In this paper, we study parallelization of $\ell$-min on many-core CPU and GPU
architectures. Although $\ell_1$-min \eqref{eq:l1min} is a convex
program, conventional algorithms such as interior-point methods
\cite{ChenS2001-SIAM,TibshiraniR1996} scale poorly
for applications such as face recognition. Recently, a number of
accelerated algorithms have been proposed that explicitly take advantage of
the special structure of $\ell_1$-min problems
\cite{LorisI2009,YangA2010-ICIP}. We investigate parallelization of a
state-of-the-art $\ell_1$-min solution based on
\emph{augmented Lagrangian methods} (ALM) \cite{BertsekasD2003,YangA2010-ICIP}.
In addition, the accuracy of the solution given by \eqref{eq:l1min_denoise} relies on the assumption
that the query image is well aligned with the training images.
In \cite{WagnerA2009-CVPR}, SBC was extended to align
a query image to each subject class individually. The alignment algorithm solves
the following problem:
\begin{equation}
\hat{\tau}_i = \arg\min_{\xx, \ee, \tau_i} \|\ee\|_1\quad \mbox{subj. to}\quad \tilde\bb\circ\tau_i = A_i\xx + \ee,
\label{eq:l1min_alignment}
\end{equation}
where $\tau_i\in T$ is in a finite-dimensional group $T$ of transformations
(e.g., affine or homography) acting on the image domain. The sequence
$\tau_i$ converges to a transformation that aligns the test image $\tilde\bb$
with the training images from the $i$-th class $A_i$. In other words, the algorithm
extends Lucas-Kanade iterative alignment \cite{LucasB1981} to use the
$\ell_1$-norm as a robust error function, while simultaneously estimating
the coefficients of the linear illumination model $A_i\x$.
Although ALM has been applied to improve the speed of a prototype face recognition system \cite{WagnerA2011-PAMI},
due to the high per-class cost
of the alignment step, the system still fails to achieve \emph{real-time}
performance against datasets of hundreds or thousands of subjects. In this paper,
in addition to accelerating the generic
$\ell_1$-min objectives \eqref{eq:l1min} and \eqref{eq:l1min_denoise}, we
also discuss how to efficiently accelerate the face alignment step
\eqref{eq:l1min_alignment} on multi-core CPUs and GPUs.
To clarify discussion, we clearly distinguish between {\em parallelism},
which is a property of an algorithm, and {\em concurrency}, which is a
property of a hardware architecture. Parallelism in an algorithm provides the
opportunity to perform computations concurrently on hardware. Algorithms often
exhibit multiple levels of parallelism, and hardware often provides multiple
levels of concurrency. In these terms, the primary contribution of this paper
will be to determine the optimal mappings between the parallelism available in
$\ell_1$-min and the concurrency available in multi-core CPU and GPU
architectures.
Finally, we present extensive benchmarks to compare the performance of the
generic ALM $\ell_1$-min algorithm and the corresponding face recognition
routines on massively parallel CPUs and GPUs. To this end, a face
recognition system for security and access control applications has been implemented,
and is diagrammed in Figure \ref{fig:pipeline}. This pipeline follows the work of \cite{WagnerA2011-PAMI}, but
the fast implementations presented in this paper are novel.
Since the parallelization of any algorithm
is highly dependent on the target hardware, we will give a brief overview of
the architecture of our test system in detail in Section
\ref{sec:concurrency}.
Finally, note that
while this paper focuses on the parallelism available in a single (shared memory) machine,
applications targeting thousands of subjects or more will likely require computation to be
distributed over a cluster or network of computers. These systems are not addressed here.
The source code of our implementations, which target NVIDIA Fermi GPUs
as well as Intel x86-64 CPUs, will be made available for peer evaluation.
\begin{figure}
\centering
{\tiny \input{figures/pipeline_simplified.tex}}
\caption{The face recognition pipeline.}
\label{fig:pipeline}
\end{figure}
\subsection{Literature Review}
Traditionally, $\ell_1$-min (a.k.a.
basis pursuit (BP)) has been formulated as a linear program
\cite{ChenS2001-SIAM}. Several variations of the solution are also well known
in optimization, including a noisy approximation via quadratic programming
called the LASSO \cite{TibshiraniR1996} and truncated Newton interior-point
method (TNIPM) \cite{KimS2007}.
One of the drawbacks of most interior-point methods for $\ell_1$-min is that they require the solution sequence to follow an
interior path via gradient descent or conjugate gradient methods, which are computationally expensive.
To mitigate these issues, an approach called \emph{Homotopy} has been recently studied to accelerate the
speed of $\ell_1$-min \cite{OsborneM2000,EfronB2004,MalioutovD2005,DonohoD2006}.
Although Homotopy can be shown to exactly estimate BP when the solution is
sufficiently sparse \cite{DonohoD2006}, the algorithm still involves
computationally expensive operations such as matrix-matrix multiplication and
linear least-squares problems with varying $A$
matrices.
%Homotopy methods for recovering sparse signals were
%first studied in the context of LASSO \cite{OsborneM2000}, which inspired a
%solution to the \emph{forward stagewise linear regression} problem called LARS
%\cite{EfronB2004}, and eventually led to the corresponding Homotopy algorithms
%for BP in \cite{MalioutovD2005,DonohoD2006}.
%\footnote{During the face alignment algorithm presented later, The $A$ matrix changes between alignment iterations but not within the L1 solver}
In Section \ref{sec:ALM}, we contend that ALM is a better choice for
implementation on many-core CPUs and GPUs. The ALM algorithm belongs to a
category of approximate $\ell_1$-min solutions called \emph{iterative
shrinkage-thresholding} (IST) methods \cite{WrightS2008,BeckA2009}. IST
algorithms mainly involve elementary operations such as vector algebra and
matrix-vector multiplication. Therefore, when the dimension of the problem
becomes high, IST-type algorithms are particularly suitable for hardware
systems with a high degree of concurrency. In \cite{YangA2010-ICIP}, the
authors showed that ALM is able to significantly improve the solver speed,
while achieving estimation accuracy competitive with other $\ell_1$-min
solutions. Therefore, we choose ALM as the core algorithm for
implementation of $\ell_1$-min in the parallel face recognition pipeline.
In terms of the past works in parallel $\ell_1$-min, the literature has been
limited, to the best of our knowledge. In \cite{BorghiA2010}, Borghi et al.
developed a special proximal gradient $\ell_1$-min algorithm based on
Moreau-Yosida regularization. In \cite{MurphyM2010}, Murphy et al. presented
parallel implementation of the l1-SPIRIT MRI reconstruction algorithm on
the same GPU architecture addressed in this paper.
\section{Augmented Lagrangian Method}
\label{sec:ALM}
In this section, we briefly describe the ALM algorithm for $\ell_1$-min
\eqref{eq:l1min} \cite{YangA2010-ICIP} and analyze its complexity. Lagrange
multipliers have been frequently used in convex programming to eliminate
equality constraints via adding a penalty term to the cost function for
infeasible points. ALM methods differ from other penalty-based approaches by
simultaneously estimating the optimal solution and Lagrange multipliers in an
iterative fashion. For $\ell_1$-min \eqref{eq:l1min}, the augmented Lagrange
function is defined as:
\begin{equation} L_\mu(\xx,\yy) = \|\xx\|_1 +
\yy^T(\bb - A\xx) + \frac{\mu}{2} \| \bb-A\xx \|_2^2,
\end{equation}
where $\mu > 0$ is a constant that penalizes infeasibility and $\yy$ is a
vector of Lagrange multipliers.
In Lagrange Multiplier Theory \cite{BertsekasD2003}, if there exists a multiplier $\y^*$ that
satisfies the second-order sufficiency conditions for optimality, then for a sufficiently large $\mu$, the optimal
$\ell_1$-min solution also minimizes
\begin{equation}
\xx^* = \arg \min L_\mu(\xx,\yy^*).
\label{eq:optimal-ALM}
\end{equation}
In practice, the optimal values for the triplet $(\xx^*, \yy^*, \mu)$ are all
unknown. Furthermore, it has been observed that solving
\eqref{eq:optimal-ALM} with a large initial value of $\mu$ tends to lead to
slower convergence speed \cite{WrightS2008,YangA2010-ICIP}. In
\cite{BertsekasD2003,YangJ2009}, an alternating procedure has been shown to
iteratively update $\xx$ and $\yy$:
\begin{equation}
\left \{
\begin{array}{lll}
\xx_{k+1} & = & \arg\min_{\xx} \, L_{\mu_k} (\xx,\yy_k)\\
\yy_{k+1} & = & \yy_k + \mu_k (\bb - A\xx_{k+1}) \\
\end{array}
\right . ,
\label{eq:alm}
\end{equation}
where $\mu_{k}\rightarrow \infty$ increases monotonically.
The iteration terminates when the estimates $(\xx_k, \yy_k)$ converge.
Note that in the iterative procedure \eqref{eq:alm}, the second
step only involves vector algebra and matrix-vector multiplication. Therefore,
the procedure is computationally efficient only if it is easier to minimize the
augmented Lagrangian $L_{\mu_k} (\xx,\yy_k)$ compared to solving the original problem
\eqref{eq:l1min} directly. In fact, this problem can be solved element-wise
iteratively by a soft-thresholding algorithm \cite{WrightS2008,BeckA2009},
whose time complexity is bounded by $O(n^2)$ and can be easily parallelized.
Algorithm \ref{alg:alm} summarizes the generic ALM $\ell_1$-min algorithm. \footnote{For conciseness, we
only present the ALM algorithm in the primal domain. There also
exist implementations in the dual domain \cite{YangJ2009,YangA2010-ICIP}.}
The soft-thresholding algorithm involves the element-wise operator $\shrink(x) = \sign(x)\max(|x|,0)$.
\begin{algorithm}[h]
\caption{Augmented Lagrangian Method (ALM)}
\small
{\bf INPUT:} $\bb \in \Re^m$, $A=[A_1,\cdots, A_K] \in \Re^{m \times n}$, $\tau\leftarrow \max\mbox{eig}(A^TA)$, and constant $\rho>1$.
\begin{algorithmic}[1]
\WHILE{not converged ($k = 1,2,\ldots$)}
\STATE $t_1 \leftarrow 1$, $\zz_1 \leftarrow \xx_k$, $\uu_1 \leftarrow \xx_k$
\WHILE{not converged ($l = 1,2,\ldots$)}
\STATE $\uu_{l+1} \leftarrow \shrink(\zz_l - \frac{1}{\tau}A^T(A\zz_l - \bb - \frac{1}{\mu_k}\yy_k), \frac{1}{\mu_k\tau})$
\STATE $t_{l+1} \leftarrow \frac{1}{2}( 1 + \sqrt{1+4t_l^2})$
\STATE $\zz_{l+1} \leftarrow \uu_{l+1}+ \frac{t_l - 1}{t_{l+1}}(\uu_{l+1} - \uu_l)$
\ENDWHILE
\STATE $\xx_{k+1} \leftarrow \uu_{l+1}$
\STATE $\yy_{k+1} \leftarrow \yy_k + \mu_k (\bb - A\xx_{k+1})$
\STATE $\mu_{k+1} \leftarrow \rho\cdot\mu_k$
\ENDWHILE
\end{algorithmic}
{\bf OUTPUT:} $\xx^* \leftarrow \xx_k$.
\label{alg:alm}
\end{algorithm}
\section{Hardware Concurrency} \label{sec:concurrency}
In this section we discuss the levels of concurrency available in the hardware
architectures considered in this paper, as well as other aspects of the
hardware that are important for performance. In particular, since caches
(regions of on-chip memory) are often orders of magnitude faster than off-chip
memory, their sizes and speeds have a dramatic effect on performance. We give a
brief overview of the caches that are available in our target architectures,
and defer discussion of their performance effects to Sections
\ref{sec:alignment_implementation_cpu} and
\ref{sec:alignment_implementation_gpu}.
Our discussion and experiments will address the most common hardware
configuration for engineering workstations: a motherboard with two quad-core
processors and a PCIe card with a single high-end GPU.
Recognition involving more than a few hundred subjects with contemporary hardware
would require a server (or cluster) configuration with an expandable
number of processors. We will not address efficient parallelization for
these systems, which may have additional challenges associated with
their non-shared memory model.
%\footnote{Non-shared memory configurations include CPU blade servers, multi-GPU servers, and clusters}.
Unless otherwise specified, all implementations utilize single precision
floating point datatypes.
\subsection{CPU Hardware Concurrency}
\label{sec:CPU-concurrency}
The main defining characteristics of contemporary multi-core CPU architectures
are that they have two levels of concurrency, relatively large amounts of
cache, and relatively high clock speeds. The baseline architecture for our experiments
is a Linux workstation with two
quad-core Intel Nehalem E5530 processors clocked at 2.4 GHz. Each processor
has its own memory interface, and is directly interfaced to half of the RAM
installed in the machine. The amount of RAM exceeds the amount used
by the algorithms, and is not an important performance consideration.
\begin{figure}
\centering
\subfigure[The larger algorithm data structures]
{
\includegraphics[scale=1.2]{figures/arrays}
\label{fig:arrays}
}
\subfigure[The caches on a E5530 CPU]
{
\includegraphics[scale=1.2]{figures/cpu_caches}
\label{fig:cpu_caches}
}
\subfigure[The caches on a GTX480 GPU]
{
\includegraphics[scale=1.2]{figures/gpu_caches}
\label{fig:gpu_caches}
}
\label{fig:caches}
\caption{A visual comparison of the algorithm working set to the CPU and GPU
caches. Note: Although the aspect ratio could not be preserved, all arrays and
caches are drawn so that the area is proportional to the size of the data. }
\end{figure}
% CPU CACHE DISCUSSION
%2MB = 2**21 bytes
%64*64*4 = 2**14 bytes = 16384 bytes per image
%... so can fit 2**7 = 128 images.
Each core has a private 32\,KiB L1 data cache and a private 256\,KiB L2 cache.
%16MB = 2**24 bytes = 16777216 bytes
%64*64*4 = 2**14 bytes = 16384 bytes per image
%... so can fit 2**10 = 1024 images.
Each processor further has 8\,MiB of L3 cache that is shared by the four cores.
Overall, the algorithm has approximately 16\,MiB of L3 cache
available for a dual-processor configuration.
For floating-point instructions, each core also has a vector processing unit
(SSE) capable of performing the same arithmetic operation on four single-precision
floating point values simultaneously.
There are thus two important levels of concurrency that need to be exploited to
efficiently use a modern CPU: {\em core-level} concurrency and {\em SSE-level}
concurrency.
\subsection{GPU Hardware Concurrency}
The main defining characteristics of contemporary multi-core GPU architectures
are that they have two (much wider) levels of concurrency, relatively small amounts of
cache, and relatively low clock speeds. Whereas most of the transistors on
a typical CPU are dedicated either to cache or to hardware that enables higher
clock speeds (such as branch prediction, out-of-order instruction execution, etc...),
most of the transistors on a GPU are dedicated to arithmetic logic units.
For our GPU implementations, we target NVIDIA Fermi GPU architecture (\eg, the GTX 480
GPU used in this paper). An explanation of the GPU programming model (CUDA) at a useful level of completeness would take more space than is available here, so we will instead frame our discussion in terms of
hardware capabilities.
%\footnote{A discussion of the CUDA programming model can be found in the NVIDIA C for CUDA Programming Guide}
CUDA programmable GPUs are comprised of several
\emph{streaming multiprocessors} (SM), each of which is roughly analogous to a CPU core.
For Fermi architecture GPUs there are up to 16 SMs, and each SM is capable of executing up to 64 single precision
floating point operations
concurrently. \footnote{In CUDA terms, the warp width is 32, and the floating pipeline can
issue up to two warps simultaneously}
% GPU CACHE DISCUSSION
% for L2 cache
% 768KB = 786432 bytes
% 64*64*4 = 2**14 bytes = 16384 bytes per image
% 48 images
Each SM contains its own L1-cache, which is divided between hardware managed
and software managed (``shared") memory. Additionally, all SMs share a common
L2-cache. For our system, each SM has 64 KiB of L1 cache, and all SMs share
768KB of L2 cache.
%The cache hierarchy
The relatively small amount
of cache (1/23 as much as CPU L3) on the GPU is balanced by a significantly
higher bandwidth between the processor chip and off-chip memory (DRAM) compared
to the CPU. A scale drawing of the caches available on the GPU can be seen in Figure \ref{fig:gpu_caches}.
The GPU has its own memory system, and any data the GPU uses must first be
transferred from CPU DRAM to GPU DRAM over PCIe. For our application,
this transfer overhead can be amortized over a large amount of computation and is
not a major concern.
While the programming model for
the SIMD units of each SM is somewhat more flexible than on the CPU,
leveraging the flexibility typically comes at the cost of reduced concurrency.\footnote{
In CUDA terms, code branches that cause warp divergence usually
result in serialization of the different groups of threads} Therefore, for
the purposes of comparing hardware architectures, the SIMD units on the GPU are
roughly analogous to the CPU SIMD units. Thus, in summary, the GPU hardware
provides two levels of concurrency: {\em SM-level}
concurrency and {\em thread-level} concurrency. Note that the GPU provides more
concurrency than a CPU at both levels (14 SMs vs. 8 cores) and (64 wide vs. 4 wide SIMD
units).
\section{Parallelism in the Face Alignment Stage}
\label{sec:alignment}
We will next discuss the parallelism available in the face recognition
pipeline, and propose techniques for exploiting this parallelism on the GPU and CPU hardware
described in the previous section. In this section we first focus on the
face alignment stage (see Figure \ref{fig:pipeline}).
{\bf
Face alignment \eqref{eq:l1min_alignment} estimates an image transformation
$\tau$ that maps pixels in a canonical frame, where the face is well aligned,
back to pixels in the original image.
This is illustrated in Figure \ref{fig:transformation}.
\begin{figure}
\centering
\includegraphics[width=3.4in]{figures/transformation}
\caption{The simililarity transformation $\tau$ maps pixels in a canonical frame
where the image is aligned back to pixel locations in the original-sized images.}
\label{fig:transformation}
\end{figure}
The training images are prepared as
follows: canonical outer-eye corner locations are defined in the canonical
frame, and the outer-eye corners are manually selected in the training images.
The training images are warped/resampled by a similarity transformation
(rotation, translation, and scale) between the two reference frames that is
computed from the eye corners.
The canonical eye corners are also illustrated in Figure \ref{fig:transformation}
During iterative alignment, the transformation $\tau$ between the original test
image and the canonical frame is initialized via a face detector, and then updated
iteratively. A different $\tau_i$ is computed separately for each training class $A_i$.
$\ell_1$ minimization is performed on canonical-sized images that
have been stacked into the columns of matrices.
Directly solving \eqref{eq:l1min_alignment} in its original formulation is difficult since
it is a non-convex problem and may exhibit local minima.
Minimization is therefore performed iteratively, alternating between updating
a linearized version of the problem at each $j$th step, and solving the linearized problem:
\begin{equation}
\min_{\xx, \ee, \Delta \tau_j}\|\ee\|_1\quad \subj\quad \tilde\bb\circ\tau_j + J_j\Delta \tau_j = A_i\xx + \ee,
\end{equation}
where $J_j = \nabla_{\tau_j}(\tilde\bb\circ\tau_j)$ is the Jacobian, and
$\Delta \tau$ is a step update to $\tau$. Denote $\bb_j =
\tilde\bb\circ\tau_j$, $B_j = [A_i, -J_j]$ and $\ww^T = [\xx^T, \Delta
\tau^T]$, then the update $\Delta \tau$ can be computed by solving the
following linear program:
\begin{equation}
\min_{\ww, \ee}\|\ee\|_1\quad \subj\quad \bb_j = B_j\ww + \ee.
\label{eq:alignment-linearization}
\end{equation}
}
%If the warping $\tau$ is constrained to be a similarity transform,
%it can be paramaterized in such a way that the computation of $\frac{\partial \bb}{\partial \tau_p}$
%reduces to several linear vector operations involving pixel coordinates $(u,v)$, $\tau \in \Re^4$,
%and the image gradients $\f_x, \f_y \in \Re^M$, the latter of which only has to be computed
%once per test image.
The per-class alignment algorithm via ALM is summarized in Algorithm
\ref{alg:iterative_alignment}.
%In terms of computational complexity, one significant
%term is the computation of the matrix pseudo-inverse $B_j^\dagger$, whose
%computational cost is bounded by $O(n_i^2m + n_i^3)$. However, for per-class face
%alignment, since both $B_j$ and $J_{j}$ are very tall matrices (\ie, $m\gg n_i$
%for all $i=1, \cdots, K$ classes), we find that the computation cost is still
%dominated by the frequently computed matrix-vector multiplications in the
%inner loop.
\begin{algorithm}[ht!]
\caption{\bf (Face Alignment via ALM)} \label{alg:iterative_alignment}
\small
{\bf Input:} $\bb$, $A_i$, $\x_0 = \mathbf{0}$, $\tau_0$, and $J_0$.
\begin{algorithmic}[1]
\WHILE{not converged ($j = 1,2,\ldots$)}
\STATE Update $\bb_j \leftarrow \frac{\bb\circ \tau_{j-1}}{\|\bb\circ \tau_{j-1}\|}$; $B_j= [A_i, -J_{j-1}]$ and corresponding $(B_j^\dagger)^T$
\STATE Initialize $\ww_0 = \mathbf 0$, $\yy_0 = \mathbf 0$
\WHILE{not converged ($k = 1,2,\ldots$)}
\STATE $\uu_0\leftarrow \ww_{k-1}$; $\zz_0\leftarrow \ee_{k-1}$
\WHILE{not converged ($l = 1,2,\ldots$)}
\STATE $\zz_l \leftarrow \shrink\left(\bb_j - B_j\uu_{l-1} + \frac{\yy_{k}}{\mu_{k-1}}, \frac{1}{\mu_{k-1}}\right)$
\STATE $\uu_l \leftarrow B_j^\dagger \left(\bb_j - \zz_{l} + \frac{\yy_{k-1}}{\mu_{k-1}} \right) $
\ENDWHILE
\STATE $\ww_k \leftarrow \uu_l$; $\ee_k \leftarrow \zz_l$
\STATE $\yy_{k} \leftarrow \yy_{k-1} + \mu_{k-1} (\bb_j - B_j\ww_{k} - \e_{k})$
\STATE $\mu_{k} \leftarrow \rho\mu_{k-1}$
\ENDWHILE
\STATE Update $\ee_j$, $\tau_j$, and $J_j$
\ENDWHILE
\end{algorithmic}
{\bf Output:} $\tau_i^*\leftarrow \tau_j, \e_i^*\leftarrow \ee_j$
\end{algorithm}
In summary, the alignment stage essentially contains two levels of available
parallelism. At the higher level, there are per-class alignment problems
that are solved independently, or \emph{problem-level parallelism}. At a lower
level, the first-order linear algebraic operations exhibit parallelism within
image operations, \ie, at the pixel level. We call this \emph{pixel-level
parallelism}. The following two sections discuss methods for exploiting
these two levels of parallelism on CPU and GPU architectures.
\subsection{CPU Implementation}
\label{sec:alignment_implementation_cpu}
Optimal implementation of Algorithm \ref{alg:iterative_alignment} on a multi-core CPU must take
into account the properties of the cache hierarchy. In general, cache that is
closer to the core (i.e. L1 cache) has higher bandwidth, but smaller size compared to cache
that is farther from the core (L3 cache). For reference, Figure \ref{fig:cpu_caches}
shows the the sizes of L2 and L3 caches of the Intel E5530 CPU,
which is a representative example of a modern multi-core CPU.
For the E5530, the L2 cache in each core is only able to store about 16 images.
For image alignment problem, the number of training images per class is
typically larger than 16.
In this paper, we compare two mappings of the parallelism available in the alignment
stage to the concurrency available on the CPU: a naive implementation that is purely
based on stock libraries and compilers, and a manually optimized version we advocate.
The naive implementation disregards
problem-level parallelism, and instead maps pixel-level parallelism
onto both the core-level and the SSE-level concurrency provided by the CPU. In
other words, alignment of the query image is performed
against a single subject at a time using all available cores on
the processors. The potential advantage of this implementation is that all of the variables
for the inner loop, most notably $B_j$ and $B_j^\dagger$, fit in L2 cache.
In contrast, our manually optimized implementation maps problem-level parallelism onto core-level
concurrency, and maps pixel-level parallelism onto SSE-level concurrency. In
other words, eight instances of \eqref{eq:alignment-linearization} are executed
in parallel with each problem running on one of the eight cores. The advantage
of this implementation is that since the cores are operating on different
alignment problem instances, no data is shared between cores, and therefore no
synchronization is necessary between the cores. Furthermore, even with eight
problem instances running concurrently, the local variables for the inner
loop still fit in L3 cache. Because of the sequential data access patterns of
the solver, the CPU hardware is able to use the full L3 cache bandwidth, which is
high enough to make the program CPU limited, rather than memory limited.
For these reasons, this implementation outperforms the previous naive solution.
%The implementation of some
%previous works \cite{WagnerA2009-CVPR,WagnerA2011-PAMI} are indeed based on the latter
%parallelization strategy.
%64 byte cache line-size per clock. L3 cache running at lower clock than cpu. Two reads and a write per clock.
%4 floats at a time in SSE.
In both implementations, most of the operations take
advantage of Intel Math Kernel Library (MKL), a commercial implementation of
the standard Basic Linear Algebra Subprograms (BLAS). For the image resampling
step, we make use of the Intel Integrated Performance Primitives (IPP) library.
Both MKL and IPP are optimized for Intel multi-core CPUs, and are able to
automatically utilize both core-level and vector-level concurrency (for the first implementation),
but are also available in single-threaded versions (for the second implementation). For
operations that are not optimized by Intel in-house libraries, such as the soft thresholding
operator in step 7 of Algorithm \ref{alg:iterative_alignment}, we achieve
vector-level concurrency via
the automatic vectorization facilities of the Intel C++ compiler (ICC) \cite{dulong1999overview}.
To achieve core-level concurrency we make use of the Open MP API. \cite{dagum2002openmp}
%We shall compare the performance of the two parallelism methods in Section \ref{sec:experiment}.
\subsection{GPU Implementation}
\label{sec:alignment_implementation_gpu}
While on the CPU, algorithm performance is highly dependent on effective
cache usage, cache is relatively unimportant on the GPU for ALM based
$\ell_1$-min. As can be seen in Figure \ref{fig:gpu_caches}, the GPU has
a very small amount of cache compared to the CPU. While it might be possible
to fit a single instance of the alignment problem (with a slightly reduced problem size)
into L2 cache, this would not be an efficient use of the GPU's resources.
The strength of the GPU's memory architecture for our purposes is its ability to sustain a very
high bandwidth to DRAM. This bandwidth is achieved by having a very large number of
threads issuing interleaved memory requests. Solving many alignment problem instances
concurrently increases the number of threads that can be used.
Therefore, our recommended GPU implementation is strikingly similar in spirit
to our recommended CPU implementation: on the CPU we use the cores to solve
multiple instances concurrently, and on the GPU we use multiple SMs for the
same purpose.\footnote{In CUDA terms, our proposed alignment stage
implementation consists of a single kernel that performs alignment for all
subject classes. Each instance of the alignment problem is assigned its own
thread block, and the GPU hardware schedules as many thread blocks to run
concurrently as the hardware will allow.} The number of subject classes that
are actually scheduled to run concurrently is chosen by the GPU hardware,
but can be indirectly influenced by manual tuning of the code (i.e., the kernel launch
configuration in CUDA terms). We have empirically determined
that performance is highest with 5-7 subject classes aligned simultaneously on each SM.
%This mapping of the problem-level parallelism onto the SM-level
%concurrency is illustrated in Figure \ref{fig:alignment_mapping_gpu}.
%\begin{figure}
%\centering
%\includegraphics[width=3.4in]{figures/alignment_mapping_gpu}
%\caption{Proposed mapping of alignment parallelism onto GPU concurrency}
%\label{fig:alignment_mapping_gpu}
%\end{figure}
Several other aspects of the implementation merit discussion.
First of all, we take advantage of the GPU's special hardware dedicated to bilinear interpolation
for the computation of $b(\tau)$ and $J(\tau)$, which essentially consist of resampled
versions of the test image and its derivatives.
Second, since there are no standard CUDA libraries that work at the SM level, we use a
custom routine for computing $B_j^\dagger = (B_j^TB_j)^{-1} B^T = G^{-1} B_j^T$, with $G$
inverted using Gaussian elimination with partial pivoting. To achieve
precision comparable to the single precision LAPACK routines in Intel's MKL
with this simplistic algorithm, we use double precision. Since $G$ is only $n
\times n$ and $B_j^\dagger$ is only computed once per optimization problem, the
cost of the inversion is dwarfed by the cost of other steps.
Similarly, sums and dot-products of large vectors are also
performed in double precision.
%The lack of a atomic floating point addition on the GPU
%significantly reduces the perfomance of this step of the algorithm.
%In Section \ref{sec:experiment}, we show the performance of our CPU and GPU
%implementations running on synthetic and real image data.
\section{Parallelization of the Face Recognition Stage}
\label{sec:recognition}
After the face alignment stage is complete, the 20 subject classes with lowest
alignment error are selected for recognition, $\bb$ and $A_i$ are re-sampled to a common alignment
using $\tau_i$, and $A_i$ are concatenated into a new matrix $A$.
A sparse representation of $\bb$ w.r.t $A$ is then recovered
by solving a single $\ell_1$-min problem, as shown
in \eqref{eq:l1min_denoise}, using Algorithm \ref{alg:alm_rec}. The
coefficients in $\x$ are then used to compute error residuals that are used for
classification.
%This recognition algorithm has
%been extensively studied in
%\cite{WrightJ2009-PAMI,YangA2010-ICIP,WagnerA2011-PAMI}.
In this stage of the algorithm, there is no problem-level
parallelism to exploit, so this section will discuss how to exploit pixel-level parallelism
on both CPU and GPU hardware.
Then in Section \ref{sec:experiment}, we will benchmark the
performance of the two architectures and further demonstrate the speed
gains achieved by our proposed implementations over previously published implementations.
\begin{algorithm}[t]
\caption{\bf (Face Recognition via ALM)} \label{alg:alm_rec}
\begin{algorithmic}[1]
\begin{small}
\STATE {\bf Input:} $\bb \in \Re^m$, $A \in \Re^{m \times n}$,
$\x_1 = \mathbf{0}$, $\e_1 = \bb$, $\y_1 =
\mathbf{0}$.
\WHILE{not converged ($k = 1,2,\ldots$)}
\STATE $\e_{k+1} = \shrink(\bb - A\x_k +
\frac{1}{\mu}\y_k, \frac{1}{\mu})$;
\STATE $t_1\leftarrow 1$, $\z_1 \leftarrow \x_k$, $\w_1 \leftarrow \x_k$;
\WHILE{not converged ($l = 1,2,\ldots$)}
\STATE $\w_{l+1} \leftarrow \shrink(\z_l +
\frac{1}{\gamma}A^T(\bb - A\z_l - \e_{k+1} +
\frac{1}{\mu}\y_k), \frac{1}{\mu\gamma})$;
\STATE $t_{l+1} \leftarrow \frac{1}{2}( 1 +
\sqrt{1+4t_l^2})$;
\STATE $\z_{l+1} \leftarrow \w_{l+1} + \frac{t_l - 1}{t_{l+1}}(\w_{l+1} - \w_l)$;
\ENDWHILE
\STATE $\x_{k+1} \leftarrow \w_{l}$, \; $\y_{k+1} \leftarrow \y_k + \mu (\bb - A\x_{k+1} - \e_{k+1})$;
\ENDWHILE \STATE
{\bf Output:} $\x^* \leftarrow \x_k, \e^* \leftarrow \e_k$.
\end{small}
\end{algorithmic}
\end{algorithm}
\subsection{Recognition Stage Implementation}
Compared to problem-level
parallelism, the exclusively pixel-level parallelism in Algorithm \ref{alg:alm_rec} is
relatively straightforward to exploit:
On the CPU, most of the operations map
well onto MKL BLAS calls, and operations that do not can be easily
parallelized using OpenMP and auto-vectorization.
On the GPU, most of the operations map well onto similar calls in NVIDIA's GPU
BLAS library (CUBLAS) which, like MKL, can take advantage of both levels of
concurrency available in the hardware architecture. Operations that do not map
well onto the CUBLAS API were implemented directly in CUDA code.
Additionally, some operations that could have been implemented via multiple
BLAS calls, performance improvements were achieved by combining multiple
vector-vector operations into a single kernel, due to reduced bandwidth and
kernel call overhead.
In order to avoid expensive data transfer across the bus connecting the GPU
card and the CPU motherboard (the PCI express bus), all of the data is
transferred to GPU DRAM once, and all non-trivial tasks are performed on the
GPU on data stored in GPU DRAM.
%with the sole exception of the pseudo-inverse of $B_j$,
%which is computed on the CPU via Intel's LAPACK implementation, and then
%uploaded to GPU DRAM.
\section{Experiments}
\label{sec:experiment}
In this section, we benchmark the
performance of our parallel implementations of $\ell_1$-min on CPU and GPU
platforms. In order to show how our $\ell_1$-min algorithms scale with problem
size, in Section \ref{sec:simulation} we will begin with benchmarks for the
general $\ell_1$-min problem on synthetic data. We will then progress in
Section \ref{sec:benchmark} to demonstrating the speed and accuracy of our
implementations as applied to the alignment and recognition stages of the face
recognition problem.
%To generate a meaningful comparison between different hardware architectures,
%in all cases we compare the performance of hardware architectures on a
%per-board basis: For CPU implementations, the benchmark makes use of all of
%the cores in as many CPU's are present. For GPU implementations, the benchmark
%makes the best use of the entire GPU chip (most GPU boards have a single GPU
%chip).
\subsection{Simulations on Random Data}
\label{sec:simulation}
The first experiment compares the performance of our proposed CPU and GPU
implementations of the general $\ell_1$-min solver (Algorithm \eqref{alg:alm}).
The $m \times n$ measurement matrix $A$ is a random Gaussian matrix, with each
entry generated from the standard normal distribution and normalized to unit
column norm. The ratio of $m/n$ is fixed at $1/2$ with $n$ varying from 1000 to 8000.
The ground truth signal, $\xx_0$ has a sparsity rate of 10\% of $m$ with
elements sampled from a normal distribution and normalized to unit column
norm. Because the ground truth signal $\xx_0$ is known, the algorithm
terminates when $\|\bx-\bx_0\| < \tau$ with $\tau=10^{-3}$. The measurement
vector is generated by $\bb = A \xx_0$.
The results of this benchmark can be found in Figure \ref{fig:random_data}.
The $x$-axis represents the size of the $A$ matrix and the $y$-axis represents
the average amount of time to complete a single problem instance. The GPU
implementation tends to be faster than the CPU implementation at solving a
single large problem, whereas the CPU implementation is faster at solving a
single small problem. {\bf The change in slope of the CPU curve is caused by a
transition from the CPU being cache-bandwidth-bound to memory-bandwidth-bound.
Since there are other data structures besides $A$, The transition between the
two regimes occurs slightly before the size of $A$ reaches $2000 \times \1000
\times 4 = 8$ MB, i.e. the size of the CPU L3 cache. }
%changed
\begin{figure}
\begin{center}
\includegraphics[width=3.35in]{figures/time_vs_matrix_size_constant_tol}
\end{center}
\caption{\small Comparison of $\ell_1$-min runtime vs. dimensions of $A$ on random data. {\bf The crossover
occurs near where the size of the problem exceeds the size of the CPU L3 cache.}}
\label{fig:random_data}
\end{figure}
\subsection{Face Recognition Pipeline Benchmark}
\label{sec:benchmark}
This section presents benchmarks of the CPU and GPU implementations of the
alignment stage (i.e., Algorithm
\ref{alg:iterative_alignment}) as well as the recognition stage (i.e., Algorithm \ref{alg:alm_rec}) of the
face recognition pipeline.
First, in order to measure the impact of solving many
$\ell_1$ problems-per concurrently on the GPU, we benchmark three implementations of the
alignment $\ell_1$-min solver with $A$ of size $5120 \times 32$, and a fixed
$50$ inner loop iterations for each of $50$ outer loop iterations. The runtime
on a GTX480 GPU is averaged over a large number of trials, which are run
sequentially or concurrently depending on the implementation.
The results are shown in Table \ref{tbl:ubench}.
Our proposed parallelization of the $\ell_1$-min used in the alignment
stage is {\em eight times} faster than an implementation solving a single
problem at a time using the stock BLAS libraries.
\footnote{The streams implementation is limited significantly by a
cap on the number of concurrent streams in the current version of CUDA}
%The results are shown below:
\begin{table}[t!]
\caption{\small Benchmark of implementations of alignment $\ell_1$-min.}
\small
\centering
\begin{tabular}{|l|c|}
\hline
Sequential solver using CUBLAS & 302\,ms \\
\hline
One problem per SM using CUDA streams & 70\,ms \\
\hline
Four problems per SM using single kernel & {\bf 36\,ms} \\
\hline
\end{tabular}
\label{tbl:ubench}
\end{table}
Using an implementation motivated by the previous result, we now benchmark our
iterative alignment implementation on real face data. For experiments on face
data we use subsets of the CMU Multi-PIE Face Database. For gallery images we
use frontal images from session 1, which contains 20 images per subject taken
under different illuminations. For test images we use images from session 2.
The training images are prepared as follows: The iterative alignment stage
seeks a similarity transformation between the coordinate frame of the
full-resolution test image and a ``canonical'' coordinate frame in which images
are compared. The training images are aligned by applying a similarity
transormation that maps two manually clicked outer eye corners to fixed
locations in the canonical frame. A $64 \times 64$
pixel window in the canonical frame is used for resampling.
Figure \ref{fig:alignment_stage_runtime} shows the average run time of the CPU and GPU implementations to align
a query image against all the subject classes separately. We vary the total number of
subject classes to show how the algorithms scale. The plateaus seen in the GPU curve
result from the GPU hardware scheduling the computation of alignment
problems in concurrent batches, but the overall trend is linear as expected.
The manually threaded CPU implementation slightly outperforms the GPU implementation,
and it surpasses the \emph{naive}
library threaded implementation by a wide margin.
The new implementation can align the query
image in $\approx 40$ ms per subject,
while the fastest previously published result \cite{WagnerA2011-PAMI}
required $\approx 600$ ms seconds per subject in a similar setting.
\begin{figure}[t!]
\centering
\includegraphics[width=3.35in]{figures/alignment_runtime_graph}
\caption{Alignment stage runtime vs. size of training database.}
\label{fig:alignment_stage_runtime}
\end{figure}
The next experiment compares the speed of the GPU and CPU implementations of
the recognition stage. It was determined that keeping 20 gallery subjects is
sufficient to ensure that the correct subject is kept for the recognition stage
with 95\% probability. Since recognition failures are typically caused by a
poor alignment, we find that keeping more subjects for the recognition stage does
not necessarily improve recognition rate.
Figure \ref{fig:recognition_stage_runtime} shows the recognition stage runtime
for canonical images of size: $32\times32$, $48 \times 48$, $64 \times 64$, $96 \times
96$, and $128 \times 128$. For all image resolutions, the problem size
is sufficiently small that the CPU is significantly faster than the GPU. Note
also that for both implementations, the recognition stage takes much less time
than the alignment stage.
\begin{figure}[t!]
\centering
\includegraphics[width=3.35in]{figures/speedVsResolution}
\caption{Recognition stage runtime vs. face window resolution}
\label{fig:recognition_stage_runtime} \end{figure}
Finally, we perform an experiment verifying the recognition accuracy of the
overall pipeline. As shown in Figure \ref{fig:accuracy_vs_resolution},
at the optimal resolution, which happens to match the alignment stage
resolution, the GPU implementation reaches 95\% recognition rate, the max
achievable given the alignment selection process. For significantly lower resolutions,
the accuracy drops off significantly. The slight difference in CPU vs. GPU
accuracy may be a result of numerical precision differences in our matrix
inversion and vector reduction routines.
\begin{figure}[t!]
\centering
\includegraphics[width=3.35in]{figures/accuracyVsResolution}
\caption{Recognition rate of the full pipeline vs. face window resolution.}
\label{fig:accuracy_vs_resolution}
\end{figure}
\section{Conclusion}
We have demonstrated that on both CPU and GPU algorithms, parallelizations of
ALM that solve multiple face alignment problems concurrently are significantly
faster than implementations that rely purely on vendor-supplied BLAS libraries.
Furthermore, thanks to a combination of faster hardware and more efficient use
of the hardware, we have demonstrated dramatic improvements in recognition
speed over previously reported implementations.
As CPU manufacturers increase the number of cores and vector widths, and as GPU
manufacturers increase the amount
of cache, both architectures are rapidly converging towards an architectural balance
that increasingly favorable for $\ell_1$-min based face recognition.
% of the same pipeline. Whereas
%earlier implementations required over a second to align a test image against
%just one of the 250 subjects in Multi-PIE session 1, the proposed
%implementations can perform the same task in
%{\bf under 10 seconds}!.
%The parallelization techniques presented here make it feasible to implement an
%access control system for hundreds of subjects on a handful of purpose-build
%multi-GPU workstations.
%We argue that in addition to vendor supplied libraries
%that attempt to leverage both levels of concurrency of the GPU, there is a need
%for standard libraries that operate at the SM level of concurrency.
\bibliographystyle{abbrvnat}
\bibliography{faces-short}
\end{document}