computes the coefficient used for unwrapping the phase. The second one
is the acquisition loop, while which images are taken at regular time
steps. For each image, the phase $\theta$ of all profiles is computed
-to obtain, after unwrapping, the deflection of cantilevers.
+to obtain, after unwrapping, the deflection of
+cantilevers. Originally, this computation was also done with an
+algorithm based on spline. This article proposes a new version based
+on a least square method.
\subsection{Design goals}
\label{sec:goals}
+The main goal is to implement a computing unit to estimate the
+deflection of about $10\times10$ cantilevers, faster than the stream of
+images coming from the camera. The accuracy of results must be close
+to the maximum precision ever obtained experimentally on the
+architecture, i.e. 0.3nm. Finally, the latency between an image
+entering in the unit and the deflections must be as small as possible
+(NB : future works plan to add some control on the cantilevers).\\
+
If we put aside some hardware issues like the speed of the link
between the camera and the computation unit, the time to deserialize
pixels and to store them in memory, ... the phase computation is
obviously the bottle-neck of the whole process. For example, if we
consider the camera actually in use, an exposition time of 2.5ms for
-$1024\times 1204$ pixels seems the minimum that can be reached. For a
-$10\times 10$ cantilever array, if we neglect the time to extract
-pixels, it implies that computing the deflection of a single
+$1024\times 1204$ pixels seems the minimum that can be reached. For
+100 cantilevers, if we neglect the time to extract pixels, it implies
+that computing the deflection of a single
cantilever should take less than 25$\mu$s, thus 12.5$\mu$s by phase.\\
In fact, this timing is a very hard constraint. Let consider a very
small programm that initializes twenty million of doubles in memory
and then does 1000000 cumulated sums on 20 contiguous values
(experimental profiles have about this size). On an intel Core 2 Duo
-E6650 at 2.33GHz, this program reaches an average of 155Mflops. It
-implies that the phase computation algorithm should not take more than
-$240\times 12.5 = 1937$ floating operations. For integers, it gives
-$3000$ operations.
-
-%% to be continued ...
-
-%% � faire : timing de l'algo spline en C avec atan et tout le bordel.
-
-
+E6650 at 2.33GHz, this program reaches an average of 155Mflops.
+
+%%Itimplies that the phase computation algorithm should not take more than
+%%$155\times 12.5 = 1937$ floating operations. For integers, it gives $3000$ operations.
+
+Obviously, some cache effects and optimizations on
+huge amount of computations can drastically increase these
+performances : peak efficiency is about 2.5Gflops for the considered
+CPU. But this is not the case for phase computation that used only few
+tenth of values.\\
+
+In order to evaluate the original algorithm, we translated it in C
+language. Profiles are read from a 1Mo file, as if it was an image
+stored in a device file representing the camera. The file contains 100
+profiles of 21 pixels, equally scattered in the file. We obtained an
+average of 10.5$\mu$s by profile (including I/O accesses). It is under
+are requirements but close to the limit. In case of an occasional load
+of the system, it could be largely overtaken. A solution would be to
+use a real-time operating system but another one to search for a more
+efficient algorithm.
+
+But the main drawback is the latency of such a solution : since each
+profile must be treated one after another, the deflection of 100
+cantilevers takes about $200\times 10.5 = 2.1$ms, which is inadequate
+for an efficient control. An obvious solution is to parallelize the
+computations, for example on a GPU. Nevertheless, the cost to transfer
+profile in GPU memory and to take back results would be prohibitive
+compared to computation time. It is certainly more efficient to
+pipeline the computation. For example, supposing that 200 profiles of
+20 pixels can be pushed sequentially in the pipelined unit cadenced at
+a 100MHz (i.e. a pixel enters in the unit each 10ns), all profiles
+would be treated in $200\times 20\times 10.10^{-9} =$ 40$\mu$s plus
+the latency of the pipeline. This is about 500 times faster than
+actual results.\\
+
+For these reasons, an FPGA as the computation unit is the best choice
+to achieve the required performance. Nevertheless, passing from
+a C code to a pipelined version in VHDL is not obvious at all. As
+explained in the next section, it can even be impossible because of
+some hardware constraints specific to FPGAs.
\section{Proposed solution}
\label{sec:solus}
+Project Oscar aims to provide an hardware and software architecture to
+estimate and control the deflection of cantilevers. The hardware part
+consists in a high-speed camera, linked on an embedded board hosting
+FPGAs. By the way, the camera output stream can be pushed directly
+into the FPGA. The software part is mostly the VHDL code that
+deserializes the camera stream, extracts profile and computes the
+deflection. Before focusing on our work to implement the phase
+computation, we give some general informations about FPGAs and the
+board we use.
-\subsection{FPGA constraints}
+\subsection{FPGAs}
A field-programmable gate array (FPGA) is an integrated circuit designed to be
configured by the customer. A hardware description language (HDL) is used to
FGPAs programming is very different from classic processors programming. When
logic block are programmed and linked to performed an operation, they cannot be
-reused anymore. FPGA are cadenced slowly than classic processors but they can
-performed pipelined as well as pipelined operations. A pipeline provides a way
+reused anymore. FPGA are cadenced more slowly than classic processors but they can
+performed pipelined as well as parallel operations. A pipeline provides a way
manipulate data quickly since at each clock top to handle a new data. However,
using a pipeline consomes more logics and components since they are not
reusable, nevertheless it is probably the most efficient technique on FPGA.
manipulate new data at each clock top and using parallelism to handle
simultaneously several data streams.
-%% contraintes imposées par le FPGA : algo pipeline/parallele, pas d'op math complexe, ...
+%% parler du VHDL, synthèse et bitstream
+\subsection{The board}
+The board we use is designed by the Armadeus compagny, under the name
+SP Vision. It consists in a development board hosting a i.MX27 ARM
+processor (from Freescale). The board includes all classical
+connectors : USB, Ethernet, ... A Flash memory contains a Linux kernel
+that can be launched after booting the board via u-Boot.
+
+The processor is directly connected to a Spartan3A FPGA (from Xilinx)
+via its special interface called WEIM. The Spartan3A is itself
+connected to a Spartan6 FPGA. Thus, it is possible to develop programs
+that communicate between i.MX and Spartan6, using Spartan3 as a
+tunnel. By default, the WEIM interface provides a clock signal at
+100MHz that is connected to dedicated FPGA pins.
+
+The Spartan6 is an LX100 version. It has 15822 slices, equivalent to
+101261 logic cells. There are 268 internal block RAM of 18Kbits, and
+180 dedicated multiply-adders (named DSP48), which is largely enough
+for our project.
+
+Some I/O pins of Spartan6 are connected to two $2\times 17$ headers
+that can be used as user wants. For the project, they will be
+connected to the interface card of the camera.
\subsection{Considered algorithms}
SPL on $N = k\times M$, i.e. the number of interpolated points.
We assume that $M=20$, $nb_s=1024$, $k=4$, all possible parts are
-already in lookup tables and only arithmetic operations (+, -, *, /)
-are taken account. Translating the two algorithms in C code, we obtain
-about 400 operations for LSQ and 1340 (plus the unknown for $atan$)
-for SPL. Even if the result is largely in favor of LSQ, we can notice
-that executing the C code (compiled with \tt{-O3}) of SPL on an
-2.33GHz Core 2 Duo only takes 6.5µs in average, which is under our
-desing goals. The final decision is thus driven by the third criterion.\\
+already in lookup tables and a limited set of operations (+, -, *, /,
+<, >) is taken account. Translating the two algorithms in C code, we
+obtain about 430 operations for LSQ and 1550 (plus few tenth for
+$atan$) for SPL. This result is largely in favor of LSQ. Nevertheless,
+considering the total number of operations is not really pertinent for
+an FPGA implementation : it mainly depends on the type of operations
+and their
+ordering. The final decision is thus driven by the third criterion.\\
The Spartan 6 used in our architecture has hard constraint : it has no
-floating point units. Thus, all computations have to be done with
-integers.
+built-in floating point units. Obviously, it is possible to use some
+existing "black-boxes" for double precision operations. But they have
+a quite long latency. It is much simpler to exclusively use integers,
+with a quantization of all double precision values. Obviously, this
+quantization should not decrease too much the precision of
+results. Furthermore, it should not lead to a design with a huge
+latency because of operations that could not complete during a single
+or few clock cycles. Divisions are in this case and, moreover, they
+need an varying number of clock cycles to complete. Even
+multiplications can be a problem : DSP48 take inputs of 18 bits
+maximum. For larger multiplications, several DSP must be combined,
+increasing the latency.
+
+Nevertheless, the hardest constraint does not come from the FPGA
+characteristics but from the algorithms. Their VHDL implentation will
+be efficient only if they can be fully (or near) pipelined. By the
+way, the choice is quickly done : only a small part of SPL can be.
+Indeed, the computation of spline coefficients implies to solve a
+tridiagonal system $A.m = b$. Values in $A$ and $b$ can be computed
+from incoming pixels intensity but after, the back-solve starts with
+the lastest values, which breaks the pipeline. Moreover, SPL relies on
+interpolating far more points than profile size. Thus, the end
+of SPL works on a larger amount of data than the beginning, which
+also breaks the pipeline.
+
+LSQ has not this problem : all parts except the dichotomial search
+work on the same amount of data, i.e. the profile size. Furthermore,
+LSQ needs less operations than SPL, implying a smaller output
+latency. Consequently, it is the best candidate for phase
+computation. Nevertheless, obtaining a fully pipelined version
+supposes that operations of different parts complete in a single clock
+cycle. It is the case for simulations but it completely fails when
+mapping and routing the design on the Spartan6. By the way,
+extra-latency is generated and there must be idle times between two
+profiles entering into the pipeline.
+
+%%Before obtaining the least bitstream, the crucial question is : how to
+%%translate the C code the LSQ into VHDL ?
+
+
+%\subsection{VHDL design paradigms}
+
+\section{Experimental tests}
+\subsection{VHDL implementation}
+% - ecriture d'un code en C avec integer
+% - calcul de la taille max en bit de chaque variable en fonction de la quantization.
+% - tests de quantization : équilibre entre précision et contraintes FPGA
+% - en parallèle : simulink et VHDL à la main
+%
+\subsection{Simulation}
-\subsection{VHDL design paradigms}
+% ghdl + gtkwave
+% au mieux : une phase tous les 33 cycles, latence de 95 cycles.
+% mais routage/placement impossible.
+\subsection{Bitstream creation}
-\subsection{VHDL implementation}
+% pas fait mais prévision d'une sortie tous les 480ns avec une latence de 1120
-\section{Experimental results}
\label{sec:results}
