Section: New Results

Reconfigurable Architecture Design

Arithmetic Operators for Cryptography and Fault-Tolerance

Participants : Arnaud Tisserand, Emmanuel Casseau, Thomas Chabrier, Karim Bigou, Franck Bucheron, Jérémie Métairie, Nicolas Veyrat-Charvillon, Nicolas Estibals.

Arithmetic Operators for Fast and Secure Cryptography. Scalar recoding is popular to speed up ECC (elliptic curve cryptography) scalar multiplication: non-adjacent form, double-base number system, multi-base number system (MBNS). But fast recoding methods require pre-computations: multiples of base point or off-line conversion. In paper [42] presented at ARITH, we presented a multi-base (e.g. (2,3,5,7)) recoding method for ECC scalar multiplication based on i) a greedy algorithm starting least significant terms first, ii) cheap divisibility tests by multi-base elements and iii) fast exact divisions by multi-base elements. Multi-base terms are obtained on-the-fly using a special recoding unit which operates in parallel to curve-level operations and at very high speed. This ensures that all recoding steps are performed fast enough to schedule the next curve-level operations without interruptions. The proposed method can be fully implemented in hardware without pre-computations. We report FPGA implementation details and very good performance compared to state-of-art results. A specific version of our method allows random recodings of the scalar which can be used as a partial counter-measure against side-channel attacks. The PhD thesis defended by Thomas Chabrier [18] deals with MBNS and other types of arithmetic recodings for ECC scalar multiplication (title: "Arithmetic recodings for ECC cryptoprocessors with protections against side-channel attacks").

In the paper [67] , presented at ComPAS, we presented efficient arithmetic operators for divisibility tests and modulo operations for large operands (e.g. 160-600 bits like in cryptographic applications) and by a set of small constants such as $(2^a,3,5,7,9)$ where $1\le a\le 12$. These operators have been validated and implemented on FPGAs.

In the paper [39] presented at CHES, we described a new RNS modular inversion algorithm based on the extended Euclidean algorithm and the plus-minus trick. In our algorithm, comparisons over large RNS values are replaced by cheap computations modulo 4. Comparisons to an RNS version based on Fermat's little theorem were carried out. Comparisons to a version based on Fermat's little theorem were carried out. The number of elementary modular operations is significantly reduced: a factor 12 to 26 for multiplications and 6 to 21 for additions. Virtex 5 FPGAs implementations show that for a similar area, our plus-minus RNS modular inversion is 6 to 10 times faster. Other implementation results of RNS for ECC cryptosystems have been presented in [75] and [74] .

ECC Processor with Protections Against SCA. A dedicated processor for elliptic curve cryptography (ECC) is under development. Functional units for arithmetic operations in GF($2^m$) and GF($p$) finite fields and 160-600-bit operands have been developed for FPGA implementation. Several protection methods against side channel attacks (SCA) have been studied. The use of some number systems, especially very redundant ones, allows one to change the way some computations are performed and then their effects on side channel traces. This work is done in the PAVOIS project.

Arithmetic Operators for Fault Tolerance. In the ARDyT project, we work on computation algorithms, representations of numbers and hardware implementations of arithmetic operators with integrated fault detection (and/or fault tolerance) capabilities. The target arithmetic operators are: adders, subtracters, multipliers (and variants of multiplications by constants, square, FMA, MAC), division, square-root, approximations of the elementary functions. We study two approaches: residue codes and specific bit-level coding in some redundant number systems for fault detection/tolerance integration at the arithmetic operator/unit level. FPGA prototypes are under development.

Reconfigurable Processor Extensions Generation

Participants : Christophe Wolinski, François Charot.

Most proposed techniques for automatic instruction sets extension usually dissociate pattern selection and instruction scheduling steps. The effects of the selection on the scheduling subsequently produced by the compiler must be predicted. This approach is suitable for specialized instructions having a one-cycle duration because the prediction will be correct in this case. However, for multi-cycle instructions, a selection that does not take scheduling into account is likely to privilege instructions which will be, a posteriori, less interesting than others in particular in the case where they can be executed in parallel with the processor core. The originality of our research work is to carry out specialized instructions selection and scheduling in a single optimization step. This complex problem is modeled and solved using constraint programming techniques. This approach allows the features of the extensible processor to be taken into account with a high degree of flexibility. Different architectures models can be envisioned. This can be an extensible processor tightly coupled to a hardware extension having a minimal number of internal registers used to store intermediate results, or a VLIW-oriented extension made up of several processing units working in parallel and controlled by a specialized instruction. These techniques have been implemented in the Gecos source-to-source framework.

Novel techniques addressing the interactions between code transformation (especially loops) and instruction set extension are under study. The idea is to automatically transform the original loop nests of a program (using the polyhedral model) to select specialized and vector instructions. These new instructions may use local memories located in the hardware extension and used to store intermediates data produced at a given loop iteration. Such transformations lead to patterns whose effect is to significantly reduce the pressure on the memory of the processor. An experiment realized on the matrix multiplication (extracted from PolyBench/C, the polyhedral benchmark suite) using an Xtensa extensible and configurable processor from Tensilica shows interesting speedups. Speedup of 4.3 for the transformed code compared to the initial code for matrices of size 512x512 and speedup of 8.75 (respectively 20.15) in case of an extension allowing SIMD vector operations on vector of 4 32-bit words (respectively 16 32-bit words) are observed.

Runtime Mapping of Hardware Accelerators on the FlexTiles 3D Self-Adaptive Heterogeneous Manycore

Participants : Olivier Sentieys, Antoine Courtay, Christophe Huriaux.

FlexTiles is a 3D stacked chip with a manycore layer and a reconfigurable layer. This heterogeneity brings a high level of flexibility in adapting the architecture to the targeted application domain for performance and energy efficiency. A virtualisation layer on top of a kernel hides the heterogeneity and the complexity of the manycore and fine-tunes the mapping of an application at runtime. The virtualisation layer provides self-adaptation capabilities by dynamically relocation of application tasks to software on the manycore or to hardware on the reconfigurable area. This self-adaptation is used to optimize load balancing, power consumption, hot spots and resilience to faulty modules. The reconfigurable technology is based on a Virtual Bit-Stream (VBS) that allows dynamic relocation of accelerators just as software based on virtual binary code allows task relocation.

We have proposed a novel approach to hardware task relocation in an FPGA-based reconfigurable fabric, allowing offline design, routing, and unfinalized placement of hardware IPs and dynamic placement of the corresponding bit-streams at run-time. Our proposal relies on a custom dual-context FPGA configuration memory organization in a shift-register manner and on a dedicated bit-stream insertion controller leading to a break-through in terms of adaptive capabilities of the reconfigurable hardware. We show that using our custom shift-register organization across the configuration memory, and under some weak constraints, can greatly reduce the overhead implied by the 1-D to 2-D mapping of the shift-register onto the logic fabric. The use of partial dynamic reconfiguration in FPGA-based systems has grown in recent years as the spectrum of applications which use this feature has increased. For these systems, it is desirable to create a series of partial bitstreams which represent tasks that can be located in multiple regions in the FPGA substrate. While the transferal of homogeneous collections of lookup-table based logic blocks from region to region has been shown to be relatively straightforward, it is more difficult to transfer partial bitstreams which contain fixed function resources, such as block RAMs and DSP blocks. To do so, we explore adding enhancements to the FPGA architecture which allow for the migration of partial bitstreams including fixed resources from region to region even if these fixed function resources are not located in the same position in the region. Our approach does not require significant, time-consuming place-and-route during the migration process. We quantify the cost of inserting additional routing resources into the FPGA architecture to allow for easy migration of heterogeneous, fixed function resources. Our experiments show that this flexibility can be added for a relatively low overhead and performance penalty. As mentioned above, the Virtual Bit-Stream (VBS) is a concept of an unfinalized, pre-routed bit-stream which could be loaded almost anywhere on a custom FPGA logic fabric. Unlike classical bit-streams, the VBS is not tied to a specific location on the circuit, hence its ”virtual” qualifier. The goal is to generate a single VBS only once for each and every possible location of the logic fabric in the FPGA in a unfinished manner: the time-consuming packing, place and route steps are done offline and only local routing is done at runtime in order to ensure fast decoding time as well as low memory overhead. The VBS concept is pending for a European patent application.

Power Models of Reconfigurable Architectures

Participants : Robin Bonamy, Daniel Chillet, Olivier Sentieys.

Including a reconfigurable area in complex systems-on-chip is considered as an interesting solution to reduce the area of the global system and to support high performance. But the key challenge in the context of embedded systems is currently the power budget and the designer needs some early estimations of the power consumption of its system. Power estimation for reconfigurable systems is a difficult issue since several parameters need to be taken into account to define an accurate model.

One first parameter concerns the choice of tasks to execute and their allocation in the computing resources. Indeed, several hardware implementations of an algorithm can be obtained and exploited by the operating system for a flexible allocation of tasks to optimize energy consumption. These different hardware implementations can be obtained by varying the parallelism level, which has a direct impact on area and execution time and therefore on power and energy consumption. To highlight this point, we made several evaluations of delay, area, power, and energy impacts of loop transformations using High Level Synthesis tools. Real power measurements have been made on an FPGA platform and for different task implementations to build a model of energy consumption versus execution time.

Furthermore, we also considered the opportunity of the dynamic reconfiguration, which makes possible to partially reconfigure a specific part of the circuit while the rest of the system is running. This opportunity has two main effects on power consumption. First, thanks to the area sharing ability, the global size of the device can be reduced and the static (leakage) power consumption can thus be reduced. Secondly, it is possible to delete the configuration of a part of the device which reduces the dynamic power consumption when a task is no longer used.

We analyzed the power consumption during the dynamic reconfiguration on a Virtex 5 board. Three models of the partial and dynamic reconfiguration power consumption with different complexity/accuracy tradeoffs are extracted. These models are used in design space exploration to include impact of reconfiguration on energy consumption of a complete system. We proposed a methodology for power/energy consumption modeling and estimation in the context of heterogeneous (multi)processor(s) and dynamically reconfigurable hardware systems. We developed an algorithm to explore all task mapping possibilities for a complete application (e.g. for H264 video coding) with the aim to extract one of the best solutions with respect to the designer's constraints. This algorithm is a step ahead for defining on-line power management strategies to decide which task instances must be executed to efficiently manage the available power using dynamic partial reconfiguration. All these results are presented in the Robin Bonamy's thesis [17]

Real-time Spatio-Temporal Task Scheduling on 3D Architecture

Participants : Quang-Hai Khuat, Quang-Hoa Le, Emmanuel Casseau, Antoine Courtay, Daniel Chillet.

One of the main advantages offered by a three-dimensional system-on-chip (3D SoC) is the reduction of wire length between different blocks of a system, thus improving circuit performance and alleviating power overheads of on-chip wiring. To fully exploit this advantage, an efficient management referring to allocate temporarily the tasks at different levels of the architecture is greatly important.

In the context of 3D SoC, we have developed several spatio-temporal scheduling algorithms for 3D MultiProcessor Reconfigurable System-on-Chip (3DMPRSoC) architectures composed of a multiprocessor layer and an embedded Field Programmable Gate Array (eFPGA) layer with dynamic reconfiguration. These two layers are interconnected vertically by through-silicon vias (TSVs) ensuring tight coupling between software tasks on processors and associated hardware accelerators on the eFPGA. Our algorithms cope with task dependencies and try to allocate communicating tasks close to each other in order to reduce direct communication cost, thus reducing global communication cost.

In the 3DMPRSoC context, our algorithms favor direct communications including: i) point-to-point communication between hardware accelerators on the eFPGA, ii) communication between software tasks through the Network-on-Chip of the multiprocessor layer, and iii) communication between software task and accelerator through TSV. When a direct communication between two tasks occurs, the data are stored in a shared memory placed onto the multiprocessor layer.

Our work in [68] takes all types of communication into consideration and proposes a scheduling and placement strategy of tasks reducing the global communication cost to 17% compared with our previous algorithm based on Pfair. In this work, the eFPGA layer of the 3DMPRSoC is supposed to contain homogeneous partial reconfiguration regions (PRR) and the size of a hardware accelerator is limited by the size of a PRR. To exceed this limitation, we analyzed the Vertex-List Structure (VLS) method for relocating hardware accelerators of various sizes anywhere onto the eFPGA if resources are available. Then, we proposed VLS-BCF algorithm [49] based on VLS that allows for reducing the overall communication cost significantly – up to 24% – compared to classical methods.

Ultra-Low-Power Reconfigurable Controllers

Participants : Vivek D. Tovinakere, Olivier Sentieys, Steven Derrien.

A key concern in the design of controllers in wireless sensor network (WSN) nodes is the flexibility to execute different control tasks for managing resources, sensing and communications tasks of the node. In this paper, low-power flexible controllers for WSN nodes based on reconfigurable microtasks are presented. A microtask is a digital control unit made up of an FSM and datapath. Scalable architectures for reconfigurable FSMs along with variable precision adders in datapath are proposed for flexible controllers. Power gating as a low power technique is considered for low power operation in reconfigurable microtasks by exploiting coarse grain power gating opportunities in FSMs and adders. Gate-level models are applied to analyze energy savings in logic clusters due to power gating. Power estimation results on typical benchmark microtasks show a $2\times$ to $5\times$ improvement in energy efficiency w.r.t a microcontroller at a cost of $5\times$ when compared with a microtask implemented as an ASIC with higher NRE costs [21] .