This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
1

An Efficient SRAM-based Reconfigurable Architecture for Embedded

Processors
Sajjad Tamimi, Zahra Ebrahimi, Behnam Khaleghi, Hossein Asadi, senior Member, IEEE

Abstract—Nowadays, embedded processors are widely used in wide

range of domains from low-power to safety-critical applications. By
providing prominent features such as variant peripheral support and Reconfigures to:  LEON2 RISC-V LEON3
flexibility to partial or major design modifications, Field-Programmable
Gate Arrays (FPGAs) are commonly used to implement either an entire
Reconfig. time t1 t2 t3
embedded system or a Hardware Description Language (HDL)-based
processor, known as soft-core processor. FPGA-based designs, however,
suffer from high power consumption, large die area, and low perfor- M Switch
mance that hinders common use of soft-core processors in low-power
M

M
Matrix Soft-Core Processor
embedded systems. In this paper, we present an efficient reconfigurable M

Logic Fetch
architecture to implement soft-core embedded processors in SRAM-based M

M
Decode
FPGAs by using characteristics such as low utilization and fragmented M
Block Write Back
accessibility of comprising units. To this end, we integrate the low M

utilized functional units into efficiently designed Look-Up Table (LUT) Select
Switch
based Reconfigurable Units (RUs). To further improve the efficiency of Matrix
the proposed architecture, we used a set of efficient Configurable Hard Execution
Logics (CHLs) that implement frequent Boolean functions while the other M M

Logic
functions will still be employed by LUTs. We have evaluated effectiveness Block
of the proposed architecture by implementing the Berkeley RISC-V M M

processor and running MiBench benchmarks. We have also examined

the applicability of the proposed architecture on an alternative open- M
Switch Switch Switch Switch
GlueMatrix
Logic Matrix Matrix
source processor (i.e., LEON2) and a Digital Signal Processing (DSP) core. Matrix
Experimental results show that the proposed architecture as compared to
the conventional LUT-based soft-core processors improves area footprint, Fig. 1. Online reconfiguration to various soft-core processors
static power, energy consumption, and total execution time by 30.7%,
32.5%, 36.9%, and 6.3%, respectively. With the limitations of hard-core processors along with infeasibility
of fabricating dedicated processors due to limitations in time and/or
I. I NTRODUCTION
Embedded systems are used in broad range of applications from cost in one hand, and prominent features of FPGAs in the other hand,
indoor utensils to medical appliances and safety-critical applications. designers have ever-increasing tendency to use soft-core processors in
The hardware infrastructure of these systems typically consists of an embedded systems. Various open-source soft-core processors [7]–[12]
embedded processor as the core and different peripherals specified have been widely used for academic and commercial purposes. Such
for the target application [1], [2]. An embedded processor can be soft-cores are developed based on full description of a processor from
implemented as two commonly used platforms, i.e., either hard- or scratch using Hardware Description Language (HDL) [13], modifi-
soft-core. Hard-core processors have received great attention due to cation of pre-tested Intellectual Property (IP) cores (e.g., MicroBlaze
the high abstraction level they provide to both developers and users and Nios II), or pre-characterized cores that are developed by higher
[3], [4]. However, commercially off-the-shelf hard-core processors abstraction level tools (e.g., SPREE [14] and Chisel [15]).
cannot fulfill design requirements when processor customizations A major challenge of embedded soft-core processors is functional
such as Instruction Set Architecture (ISA) update is demanded by the units with low utilization and/or fragmented accesses, e.g., floating
customers. In addition, these processors typically provide specific and point multiplier and divider. Such units are accessed sporadically
limited peripheral support that may not fit appropriately to a variety for a short period of time and remain idle for significant amount
of embedded applications. Moreover, there is always likelihood of time. The large period of idle time can lead to significant static
of obsolescence of a hard-core processor which complicates the power dissipation which is a key design constraint in battery-powered
migration of the overlying embedded system to newer technologies mobile devices. In particular, with the downscaling of the transistor
and leads to expensive and even infeasible design update. feature size and threshold voltage in smaller technology nodes, the
An alternative approach to implement embedded processors is role of static power has become more pronounced [16], leading to the
using reconfigurable devices such as Field-Programmable Gate Ar- era of Dark Silicon [17]. Furthermore, these low utilized units suffer
rays (FPGAs) which is referred to as soft-core. A typical soft-core from larger area (i.e., transistor count), which is another testimony
processor, as illustrated in Fig. 1, can be reconfigured on an individual for the ever-increasing static power in embedded systems.
FPGA device during system runtime. The main privilege of soft-core To improve the power consumption of embedded processors,
over hard-core counterpart is its intrinsic flexibility to reconfigure previous studies have suggested to replace the low utilized units
itself with workload variations to provide higher level of parallelism with non-volatile reconfigurable units which can be reconfigured
which results in (a) shorter time-to-market, (b) inexpensive design whenever an access to an unavailable unit is issued [18], [19]. Such
update with upcoming technologies, (c) continuous adoption with ad- a design suffers from area overhead caused by replacing the static
vancements in SRAM-based FPGAs, and (d) flexibility to implement CMOS units with Look-Up Table (LUT)-based equivalents. Several
diverse range of applications. FPGAs, however, are conventionally previous studies have also considered the use of reconfigurable
equipped with abundant logic and routing resources to target general architectures in conjunction with [20]–[23] or within the FPGA-based
applications which results in high power consumption, large silicon processors [18], [19], [24], [25] mainly to speed up the computations.
die area, and low performance compared with Application Specific Nevertheless, these architectures have two main drawbacks. First, the
Integrated Circuit (ASIC) counterparts [5]. These intrinsic charac- suggested reconfigurable architectures are designed only for a specific
teristics of FPGAs are in contrary with fundamental requisites of application domain and are not generic. Second, the suggested recon-
embedded systems [6]. figurable architectures which have been used within these platforms

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
2

are power inefficient. These shortcomings obstruct the use of such efficient synthesis algorithm which exploits efficient CHLs for im-
architectures in the majority of embedded systems. plementing significant portion of functions in the proposed RUs. To
In this paper, we propose an area and power efficient reconfigurable the best of our knowledge, this is the first effort which investigates
architecture for soft-core embedded processors by exploiting the low the applicability of CHLs in either soft or hard processors.
utilization and fragmented accesses of functional units. The proposed • We investigate the generality of the proposed architecture on an
architecture is motivated by large area and high power consumption of alternative embedded processor (LEON2) as well. In addition, to
the low-utilized functional units in SRAM-based soft-core processors. examine the insensitivity of shared RU, we use both academic and
Such an architecture, however, necessitates a comprehensive profiling commercial synthesis tools.
to determine the accessibility of processor functional units across The rest of this paper is as follows. The related work is reviewed
a wide range of applications to (a) specify which functional units in Section II. The motivation behind the proposed architecture is
should be replaced with reconfigurable component and (b) determine explained in Section III. We articulate the proposed architecture
how to efficiently integrate these units. For this purpose, we first in Section IV. Details of the experimental setup and results are
profile a wide range of benchmarks to determine infrequently-used elaborated in Section V. Finally, Section VII concludes the paper.
and fragmentally-accessed units that are the major contributor of
the logic static power, and hence, are promising candidates to be
II. R ELATED W ORK
merged in an individual Reconfigurable Unit (RU). Afterwards, to
further improve the area and power efficiency of the proposed RU, Different approaches have been proposed to reduce the area and
we propose an algorithm to efficiently integrate these components as power of soft-core processors as well as to increase their performance,
shared RUs. The proposed algorithm aims to minimize the number which can be classified in two categories. Approaches in the first
of configuration cells and logic resources in the RU by exploiting category employ three techniques, (a) static or (b) dynamic power
the similarities along with the non-uniform distribution of Boolean gating of components of execution stage [27]–[29], [36], and (c) dual
functions in the components. These homomorphic-structure functions threshold voltage to reduce power. Static power gating is applied
are implemented as a set of small area-size Configurable Hard Logics offline at the configuration time. Therefore, the power efficiency
(CHLs). To our knowledge, this is the first study that, in a fine-grained of this technique is restricted to permanently-off components which
manner, improves the area and static power efficiency of FPGA-based depends on the target application. Dynamic power gating is applied
soft-core processors using the concept of spatiality in component online, i.e., during the application runtime, so it may have higher
utilization. The proposed architecture can be employed in tandem opportunity for power reduction. In practice, however, dynamic power
with other methods that intend to architecturally improve the soft- gating encounters several shortcomings that hinders practicality of
core processors, e.g., multi-threading [26], power gating [27]–[29], this technique. We will discuss advantages of the proposed architec-
or ISA customization [30], as they are orthogonal to our proposed ture over dynamic power gating method in Section V.
architecture. As the second category, using reconfigurable devices to imple-
To evaluate the efficiency of the proposed architecture, we first ment contemporary processors is proposed. Previous studies in this
execute MiBench benchmarks [31] on Berkeley open-source RISC- category can be classified as coarse-grained and fine-grained ar-
V processor [8] to obtain the access patterns and utilization rate chitectures. The former allows bit-level operations, while the latter
of the RISC-V units. Afterwards, using an in-house script, we enables relatively more complicated operations. In the coarse-grained
identified the candidate units to be integrated in an optimal number architectures [20]–[23], the reconfigurable devices are used as co-
of RUs. Berkeley ABC [32] and Altera Quartus [33] are exploited processor to speed up computations and enhance the parallelism
for synthesizing the selected units and integrating into a shared RU. degree and energy efficiency by loop unrolling. The role of com-
Thereafter, latest version of COFFE [34] is exploited to generate piling algorithms, i.e., hardware-software partitioning, considering
efficient transistor sizing used in HSPICE. Finally, we use Verilog- today’s complex many-core architectures and multi-programming
To-Routing (VTR) 7.0 [35] fed with the parameters obtained from can adversely affect the efficiency of such architectures [37]. Thus,
HSPICE and Design Compiler to place and route the processors and aforementioned architectures impose considerable area overhead of
obtain the design characteristics. Furthermore, we examine generality profiler and departure compilers from their straightforward flow
of the proposed architecture on an alternative open-source processor to support branch prediction which may not always be correct in
(LEON2), as well. Experimental results demonstrate that our pro- interactive and time-dependent embedded applications [23].
posed architecture improves the area, static power, energy, critical In fine-grained architectures, one or several types of instructions
path delay, and total execution time of RISC-V by 30.7%, 32.5%, are selected to be executed on the reconfigurable platform. The main
36.9%, 9.2%, and 6.3%, respectively, compared to the conventional challenge in such architectures is choosing computation-intensive in-
modular LUT-based soft-core processors. Examining the applicability structions. In this regard, CHIMAERA [24] has proposed integrating
of the proposed architecture on LEON2 revealed that the area, static an array of Reconfigurable Functional Units (RFUs) into the pipeline
power consumption, and energy of LEON2 is enhanced by 23.1%, of a superscalar processor which are capable of performing 9-input
23.2%, and 23.2%, respectively. functions. The authors also proposed a C compiler that attempts
The novel contributions of this work are as follows: to map the instructions into RFUs [24]. The OneChip architecture
• We present a comprehensive design space exploration among that integrates an RFU into the pipeline of a superscalar RISC
prevalent infrastructures for datapath computational units of a processor has been proposed in [25]. The RFU is implemented in
processor to discriminate the proper infrastructure for embedded parallel with Execute and Memory Access stages and includes one
systems. This suggests integration of low utilized and fragmentally- or more FPGA devices and a controller. Thus, it can be internally
accessed components into shared RUs. pipelined and/or parallelized. In addition to significant area and power
• For the first time, we propose a metric to efficient integration overheads imposed by multiple reconfiguration memories and addi-
of units in elaborately designed RUs based on component area, tional circuitries, design complexity (e.g., modifying the compiler)
utilization ratio, and configuration time. is another drawback of the OneChip architecture. A heterogeneous
• By comprehensive characterization of most frequent and architecture exploiting RFUs based on Spin-Transfer Torque Random
homomorphic-structure functions, we propose an area and power Access Memory (STT-RAM) LUTs for the hard-core general-purpose

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
3

TABLE I
S UMMARY OF EXPLOITING RECONFIGURABLE FABRICS IN PROCESSORS
Granularity Modification Structure (Topology) Improvement/Design Overhead
8x8 array of 32-bit ALU/multiplier, register file, ✓ Performance
MorphoSys [20] Coarse-grained Not mentioned
modified TinyRISC processor, and memory unit 5 No compiler support, no area and power evaluation
Developing specific VLIW processor with reconfigurable ✓ Performance 5 Large configuration overhead,
ADRES [21] Coarse-grained
compiler framework matrix and register file limited benchmarks, no area and power evaluation
✓ Reconfiguration time
Developing specific Array of Processing Elements (PEs) contains 8-bit
PipeRench [22] Coarse-grained 5 Performance (PEs communicate through global I/O),
compiler framework ALU (LUT and carry chain) and register file
large bitstream, no area and power evaluation
ARM7 microprocessor with a FPGA contains ✓ Performance and energy due to loop unrolling
Warp [23] Coarse-grained Not mentioned
32 bit MAC, loop control hardware, and data address 5 3-4X area overhead
Dynamically scheduled out of order superscalar
✓ Performance
CHIMAERA [24] Fine-grained Modified GCC compiler processor contains 32 rows of 4-LUTs, carry chain,
5 No area and power evaluation
shadow register file, cache, and control unit
Developing specific Superscalar pipeline,processor integrated,with RFU ✓ Performance
OneChip [25] Fine-grained
simulator framework (parallel to EXE and MEM stages) 5 No power evaluation
Reconfiguration and Replacement of ASIC-based low utilized functional ✓ Performance and power
[18], [19] Fine-grained
task migration algorithm units with STT-RAM 3-LUTs in IBM PowerPC 5 Area overhead
Integrating low utilized units in shared reconfigurable ✓ Area, power, performance, and energy
Proposed Fine-grained No modification required
units in embedded processors (RISC-V and LEON2) 5 Reconfiguration overhead (slight)

processors has been proposed in [18]. The proposed RFUs attempt

to mitigate unavailability of functional units using either static or dy- RISC-V Rocket Tile Integer Unit
Error Slave
namic adaptive algorithms for reconfiguration. In the static algorithm,
Mem/IO ALU
during a learning phase, idle functional units are identified, and are Converter Control & Status
Registers (CSR)
reconfigured to most active units to provide higher availability. In the IU-
Queue MUL/DIV
dynamic algorithm, the reconfiguration decision is made periodically dcache
Uncore
while at the beginning of each interval, idle units are replaced with Host Instruction Decode
the active ones. Although 18% improvement in performance has Target

Write Back
Interface
been reported, considerable area and power overhead is expected Outer
Memory Instruction Fetch
due to larger area and higher power consumption of a reconfigurable
System
functional unit compared with custom ASIC or standard cell based
Memory Page Table Word Cache Controller
components. More importantly, substantial modifications in both Controller FPU Unit Decoder FP-FP2FP

compiler and processor architecture, especially bus lines and datapath, FP-INT2FP
FP-Single
Interrupt [MUL/ADD/SUB] FP-FP2INT
are required which are neglected in this study. A similar approach Manager
along with thermal-aware reconfiguration of functional units has
icache
been proposed in [19], wherein computations of thermally hotspot DMA FP-Double
FP-DIV/SQRT
functional units are migrated to cooler units in order to balance the [MUL/ADD/SUB]

temperature distribution. Aforementioned drawbacks for [18] stand Slow IO

here, as well. The summary of previous studies is reported in Table
I. Low Utilization High Utilization Not Considered Memories

III. M OTIVATION Fig. 2. RISC-V processor (area of each block is normalized to the chip area)
The main prerequisite to integrate several functional units into a
TABLE II
single shared RU is low utilization and fragmented access of each
C OMPONENTS U TILIZATION AND A REA C OMPARISON IN RISC-V
unit. Low utilization is a crucial constraint since implementing even
few high utilized components as a single RU is inefficient as frequent Computational Units
Operation Descriptions
# 4-LUTs
(Area ratio Utilization
in Processor Core
access to such components leads to substantial execution time and without Memories)
Integer ALU Integer Addition and Subtraction 1150 (2.4%) 44.8%
power penalties. Fragmented accesses is another important criterion Unit IU-MUL/DIV Integer Multiplication and Division 4573 (9.5%) 1.8%
which facilitates integration of several functional components into a FP-Double Double Precision Multiplication,
10918 (22.9%) 0.16%
[MUL/ADD/SUB] Addition, and Subtraction
single RU. To comprehend this, assume two components C1 and C2 32/64-bit Integer Numbers to
(IEE754 2008 standard)
Floating Point Unit

FP-INT2FP Single/Double Precision Floating 741 (1.5%) 0.09%

both having 50 percent utilizations. Accordingly, two access patterns Point Format Conversion
to these components could be either as: Single/Double Precision Floating
FP-FP2INT Point Numbers to 32/64-bit 2330 (4.3%) 0.03%
(a) {C1 , C1 , C1 , C1 → C2 , C2 , C2 , C2 → C1 , C1 , C1 , C1 → Integer Format Conversion
NaN (not a number) values to
... → C2 , C2 , C2 , C2 } FP-FP2FP
an acceptable numbers Conversion
2074 (19.4%) 0.02%

(b) {C1 → C2 → C1 → C2 → ... → C1 → C2 } FP-Single

[MUL/ADD/SUB]
Single Precision Multiplication,
Addition, and Subtraction
3152 (6.58%) 0.01%

wherein an arrow (→) denotes a need to reconfiguration. While both Single/Double Precision Division
FP-DIV/SQRT 9269 (4.9%) 0.001%
and Square Root Operations
scenarios result in the same utilization rate, obviously, the former Other Fetch Instruction Fetch Unit 11140 (23.3%) -
Pipeline Decode Instruction Decode Unit 1974 (4.1%) -
requires fewer number of reconfigurations. Stages Write Back Write Back Unit 571 (1.2%) -

A. RISC-V Processor
We have examined these conditions by running MiBench bench- occupy substantial portion (more than 44%) of total processor area
marks on RISC-V processor. RISC-V is a 64-bit in-order, open- (including caches). Area of caches and memory cells are calculated
source embedded processor suited for FPGA devices [38], [39] which using CACTI (version 6.5) [40]. We modelled default RISC-V cache
consists of six pipeline stages, wherein its execution stage is equipped configuration in CACTI for both of I-cache and D-cache which both
with both integer and floating point units, as summarized in Table have 16 KB size with 16-bytes block size, 4-way associate, 32-bit
II. Fig. 2 illustrates RISC-V components proportional to their area input/output address and data bus, and tag size of 20 for I-cache and
footprint on the silicon die. As shown in this figure, functional units 22 for D-cache. Other parts of the processor such as fetch, decode,

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
4

Fig. 3. Utilization rate of RISC-V components running MiBench benchmarks

us to propose one individual RU instead of several under-utilized

units that may never be used in most applications. As an example,
Fig. 5 illustrates the fragmented accessibility of three functional units
within FPU running the FFT benchmark which has the highest SR.
As shown in this figure, for a long sequence of instructions, one of
the units in the FPU unit is busy while the other two components are
idle. This motivates the concept of shared reconfigurable unit from
two perspectives. First, since at most one of the internal units in
the FPU unit is accessed each time, integrating these internal units
is conceivable. Second, the large idle time of unused units imposes
Fig. 4. Average inter-component switching ratio in the proposed RUs significant leakage power, particularly considering their large area.

FP-FP2INT Idle Unit IV. E FFICIENT SRAM- BASED R ECONFIGURABLE

A RCHITECTURE
FP-FP2FP Busy Unit Here, we first explore the design space for possible implementa-
500 tion of functional units inside a processor and discuss advantages
Instruction
and limitations of each design style. Afterwards, we elaborate our
FP-INT2FP Idle Unit
proposed architecture.

Instructions A. Design-Space Exploration

Fig. 5. Access pattern to low-utilized functional units in RISC-V processor Based on the underlying infrastructure, i.e., ASIC or ReConfig-
urable (RC), the implementation structures for functional units can
write-back stage, and peripheral units are illustrated by shaded blocks.
be classified in four categories which are demonstrated in Fig. 6. We
Diamond blocks are memories such as registers and caches.
will compare these design concepts against the proposed architecture
B. Resource Utilization and Access Patterns with respect to area and performance.
ASIC/module and RC/module: These architectural design styles
The utilization rate of RISC-V functional units over 3.4 × 109 (shown in Fig. 6.b and Fig. 6.d, respectively) are the conventional
instructions for MiBench benchmark is illustrated in Fig. 3. As shown architectures for implementing hard-core and soft-core processors.
in this figure, the highest accessed unit (i.e., IU-MUL/DIV) has been One major drawback of these design approaches is that the access
utilized only for 0.8% on average in all benchmarks. Therefore, the pattern to the target module has not been well-considered. As it will
first condition, i.e., low utilization, is fulfilled. It should be noted that be discussed later, there are infrequently-accessed modules within
the Arithmetic Logic Unit (ALU) which has utilization rate greater embedded processors that dedicating separate units for each of them
than 43.3% is excluded from the results. The ALU is frequently used is inefficient in terms of area and power. It has been reported
by other pipeline stages, e.g., instruction and decode stage. that area and performance gap of ASIC/module and RC/module
Fragmented accesses to the low-utilized units in processor data implementations of total processor core are 17-27x and 18-26x,
path can alleviate the reconfiguration overhead and, subsequently, the respectively. A detailed comparison of processor main building blocks
execution time. To specifically clarify the importance of fragmented with respect to area and delay has also been presented in [41].
accessibility, we use the Switching Rate (SR) parameter (i.e., number
ASIC/inst: Designing a specific module per each instruction, as
of inter-unit switching to the total instructions) between components
shown in Fig. 6.a, can significantly speed up the logic operations by
for each benchmark1 . For example, if accesses to three units are
making them dedicated, hence, small and efficient. Nevertheless, its
{C1 , C1 → C2 → C3 , C3 , C3 → C2 }, then SR is 73 where 3
main drawbacks such as complicated control unit and large address
is the number of inter-unit switching and 7 is the total number of
buses cause significant area and power, which limits performance
instructions. The results reported in Fig. 4 show that, on average,
gain, and makes it impractical to implement all instructions.
SR is 4.9 × 10−3 in MiBench benchmarks, which indicates one
1 RC/all: In this implementation, a single reconfigurable unit is ded-
switching per 204 instructions (i.e., 0.0049 ). This value would become
icated to all instructions which is represented in Fig. 6.c. This design
one switching per 535 instructions if FFT benchmark is excluded.
style incurs significant reconfiguration overhead since it requires to be
Notice that some benchmarks utilize only a unique component
reconfigured once a new instruction needs to be executed. Area (i.e.,
during the whole execution time. For example, jpeg benchmark only
the number of LUTs) in such an architecture is determined by area
utilizes MUL-DIV unit as it can be observed in Fig. 3. Thus, SR
of the largest module. Beside large execution time of applications,
for such benchmarks is zero. This observation has further motivated
the potential area efficiency of such architecture will be shaded if
1 We take the average in the benchmarks having multiple subs.
the design comprises a set of small and frequently used modules and
large but infrequently used ones.

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
5

reconfiguration time (which is typically more than 50us) will domi-

Instruction Instruction Instruction Instruction Instruction Instruction Instruction Instruction
type 1 type 2 type N type 1 type 2 type 3 type N-1 type N nate the overall execution time (can be estimated as 500×5ns = 5us
... ... assuming a clock cyle of 5ns). As a result, we can alter this formula
`
SC `
SC `
SC `
SC ` SC `
SC to λef f ∝ Area × Trcf g . In addition, the critical path delay of a
component component component component component component
#1 #2 #N #1 #2 #M circuit is not affected by the number of units inside an RU since
(a) ASIC/inst (b) ASIC/module each time only one component is implemented, and thereby, can take
its original mapping. Assuming N RUs, λef f can be obtained from
Instruction Instruction Instruction Instruction Instruction
...
Instruction Instruction Instruction
type 1 type 2 type N type 1 type 2 type 3 type N-1 type N Equation 2. In this equation, area of each RU (in term of number
... of LUTs) is equal to the area of its largest candidate unit. Thus,
Single LUT LUT LUT LUT integrating Components (denoted by Cj ) with similar area/size into
Component Component Component Component
#1 #2 #M an individual RU is more prosperous.
(c) RC/all (d) RC/module
∑
N
λef f ∝ Area(RUi ) × Trcf g
Fig. 6. Baseline architectures of implementing the functional units: (a) i=1
ASIC/inst (b) ASIC/module (c) RC/all (d) RC/module (2)
∑
N
= max{Area(Cj )|Cj ∈ RUi } × Trcf g
High-Frequent High-Frequent High-Frequent Low-Frequent Low-Frequent Low-Frequent Low-Frequent
Instruction Instruction Instruction Instruction Instruction Instruction Instruction i=1
type 1 type 2 type K-1 type K type K+1 type N-1 type N

... ... The number of reconfigurations in a particular RU can be obtained

from the original benchmark trace by discarding the non-existing
LUT LUT LUT LUT units in that RU. For instance, if the original instruction sequence is
component component component
#1 #M1 #1
component {C1 , C2 , C1 , C3 , C1 , C2 , C3 , C2 } and only C1 and C2 are integrated
#M2
inside this RU, then the number of reconfigurations, i.e., Trcf g , would
LUT-RU 1 2 3
be equal to three, (i.e., {C1 − → C2 − → C1 ,C3 , C1 −
→ C2 ,C3 , C2 }).
Fig. 7. The proposed LUT-RU architecture overview We carefully analyzed λef f for possible architectural assortments
B. Proposed Architecture of grouping six candidate components into two or three RUs which
The base of the proposed architecture is to integrate the low- has 40 different cases as summarized in Fig. 8. All λef f in this
utilized and fragmentally-accessed units into shared RUs to reduce figure are normalized to the minimum case (i.e., the case with
logic resources and the number of configuration bits in the processor. RU 1 = [1, 4], RU 2 = [2, 3], and RU 3 = [5, 6]). For the sake
Hence, this will improve area and power efficiency of the soft-core of brevity, we have omitted the results of integrating candidate
processor. To achieve an efficient architecture, we take the following units into four and five RUs, since such classification does not gain
steps: (a) finding the low-utilized candidate units to be integrated much area-power improvement. Based on the results reported in Fig.
in RUs, (b) appropriate grouping of the candidate units into single 8, we propose three RUs based on the best classification, which
or multiple RUs, and (c) optimizing the proposed architecture by is detailed in Table III. i.e., RU1=(FP Single[MUL/ADD/SUB],
exploiting the similarities in Boolean functions of associated units. FP2INT), RU2=(FP Double[MUL/ADD/SUB], FP-DIV/SQRT), and
We explain each step in the following. RU3=(FP2FP, INT2FP).
1) Candidate Units: As discussed in Sec. III, low utilized and Exploiting the concept of integrating several low-utilized and
fragmentally-accessed units (i.e., small number of switching to/from fragmentally-accessed units in one RU (LUT-RU architecture) im-
a unit) is essential to migrate them to an RU. These parameters are proves several design parameters. First, the number of reconfigurable
determined by running the target application benchmarks and probing units in the pipeline stages will be reduced, and therefore the
the trace file. For the RISC-V processor, as shown in Fig. 3, ALU complexity of the control unit and address bus will be reduced.
and IU-MUL/DIV units have relatively high utilization (44.8% and Second, number of write-back ports decreases, which consequently
1.5%, respectively) compared to the average utilization rate of the reduces the number of write-back buffers. It should be noted that
floating-point units (0.24%). On the other hand, the area of the ALU the proposed architecture would not provide considerable advantage
is smaller than the FPU. Therefore, it is not effective to migrate when (a) functional units are high utilized and have frequent access
the ALU unit into a reconfigurable unit. As presented in Table II, patterns that may cause considerable execution time overhead and
FPU unit occupies a significant fraction (more than 59.5%) of the (b) functional units differ significantly in size, e.g., merging two
processor area while its utilization is only about 0.24%. The same component with 1× and 20× LUTs may not be beneficial since the
trace is used to obtain the total number of switching between units context switching overhead can shallow the small area gain.
for each benchmark and estimate the reconfiguration overhead. C. Optimizing RUs
2) Grouping into RUs: The critical step in the proposed architec- By following the previous steps, the proposed LUT-RU architecture
ture is determining the number of RUs and assigning the candidate could be achieved. Here, we aim to further improve the efficiency of
units to each unit. There is a trade-off between the number of RUs and the proposed architecture by reducing the number of resources i.e.,
total execution time. Smaller number of RU results in area-efficient transistors and configuration cells, by using area and power efficient
architecture (i.e., more units are bounded in an RU) but increases the
overall execution time due to the high number of reconfigurations in TABLE III
G ROUPING CANDIDATE UNITS INTO PROPOSED RU S
these units. We define the efficiency goal (λef f ) as minimizing the
area × execution time according to Equation 1. RUs Area (4-LUT) Description
λef f = Area × Texecution (1) FP-Single[MUL-ADD-SUB]
RU1 3152
FP-FP2INT
In this equation, Texecution is composed of two parts, (a) required FP-Double[MUL-ADD-SUB]
time to execute the instructions and (b) reconfiguration overhead RU2 10918
FP-DIV/SQRT
(which can be the dominant part). Even in the case that reconfigura- FP-INT2FP
RU3 2074
tion is performed infrequently (e.g., once per 500 instructions), the FP-FP2FP

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
6

Fig. 8. Various architectural classification based on λef f

our investigation has revealed that CHL1 and CHL2 can implement
High-Frequent High-Frequent High-Frequent Low-Frequent Low-Frequent Low-Frequent Low-Frequent 68% of functions in OR1200 in VTR benchmark suite and 59% of
Instruction Instruction Instruction Instruction Instruction Instruction Instruction
type 1 type 2 type K-1 type K type K+1 type N-1 type N functions in TV80 benchmark in IWLS’05. It is also noteworthy
... ... that any generic function, even unidentified at design time, can be
implement using pure CHLs. Albeit, in such non-LUT designs there
LUT LUT CHL-LUT
component component
CHL-LUT will be an overhead in the number of used logic blocks (CHLs) due
component component
#1 #M1 #1 #M1 to elimination of 4-LUT and pure exploiting of CHLs which are less
CHL-LUT-RU
flexible than LUT in one-to-one implementation of some functions.
Fig. 10 illustrates the proposed Configurable Logic Block (CLB)
Fig. 9. The proposed CHL-LUT-RU architecture overview wherein cluster size (N) is equal to 10 and CHL1:CHL2:LUT ratio is
CHLs instead of conventional LUTs. Previous studies have shown equal to 4:3:3, based on observations obtained from Table IV. Each
that 4-LUT based FPGA architecture has the minimum area among proposed CLB reduces area by 28.6% compared to homogeneous
different LUTs [42]. Nonetheless, even in such a small-input LUT, LUT-based CLB. The proposed CHL-LUT-RU approach iterates over
the majority of LUT structure remains underutilized in a broad range functions of components mapped to an RU and examines whether the
of circuits [43]–[45]. For example, a considerable portion of functions function could be substituted with either one of CHLs. Otherwise, it
do not necessarily have 4-inputs, so half of LUT configuration bits will remain as a 4-LUT.
and multiplexer structure, or more, is wasted. Even if all inputs
D. Insensitivity of Proposed Architecture to Different Synthesis Tools
are used, a small set of specific functions are repeated significantly
higher than others. Therefore, there is no need to allocate generous In addition to an academic synthesis tool (Berkeley ABC), we
flexibility of 16 SRAM cells to implement them. The high number repeat the mapping flow with a commercial tool (Altera Quartus
of SRAMs in LUT-based logic blocks can exacerbates the area and Integrated Synthesis (QIS) tool [33]) to investigate the sensitivity
power consumptions in FPGA and soft-core processors. In recent of the results to CAD flow. To this regard, QUIP targeting Startix IV
studies, several fine-grained architectures have been proposed that device is exploited which directly synthesizes a VHDL, Verilog, or
use small reconfigurable logic cells solely or along with conventional SystemVerilog circuit to hierarchical BLIF. Our investigation reveals
LUTs [43]–[45]. CHLs are logic blocks that can implement the that the coverage ratio of CHLs is almost the same obtained by
majority of functions with considerably smaller area and less delay an academic synthesis tool (Berkeley ABC). The comparison of the
as compared to their LUT counterparts. However, the main goal of CHLs coverage ratio in both Berkeley ABC and Altera Quartus is
these studies is enhancing of design constraints in FPGA platforms detailed in Table IV. It is noteworthy that default synthesis constraint
while their applicability in soft-core processors is not evaluated. of ABC is area optimization while Quartus aims to reduce the delay
Accordingly, we have proposed an area-power optimization ap- which results in larger area (number of LUTs) and therefore power
proach, i.e., CHL-LUT-RU, where for each RU, we synthesize each consumption [33]. To be consistent with other studies, we evaluate
component (which is assigned to the target RU) into 4-input LUTs our proposed architecture using netlists of Berkeley ABC.
by the means of Berkeley ABC synthesis tool [32] to obtain 4-
input functions and then extract its Negation-Permutation-Negation V. E XPERIMENTAL S ETUP AND R ESULTS
(NPN) class representation. To elaborate, F = AB̄ + CD and Here, we elaborate the implementation and evaluation flow of
G = BC + AD are NPN-equivalent since each function can be the proposed and baseline architectures. According to the overall
obtained from the other by negating B and permuting A and C. There- evaluation flow illustrated in Fig. 11, we first examine the access pat-
fore, such functions can be implemented with the same homomorphic tern to each of the datapath components to identify low-utilized and
logic structure augmented with configurable inverters at its inputs TABLE IV
and output. Our investigation, reported in Table IV, has revealed C OVERAGE RATIO OF 4- INPUT NPN- CLASSES IN LOW- UTILIZED
that a significant portion (66.8%) of functions in the proposed RUs COMPONENT FUNCTIONS IN RISC-V PROCESSOR

have the same homomorphic structure based on the concept of

Cell NPN-Class Berkeley ABC Altera Quartus
NPN-class representation and can effectively be implemented with
CHL1 and CHL2 which have been primarily introduced in [44]. It ABCD
should be noted that although applicability of CHLs is motivated CHL1 AB+CD 38.4% 33.4 %
by function characterization in RISC-V processor, their effectiveness AB(C+D)
is also examined in LEON2 processor and large-scale commercial A(B+CD)
CHL2 28.4% 21.1%
FPGA benchmarks [44], [45] wherein some of these benchmarks are A(B+C+D)
soft-core processors and microprocessor by themselves. For instance, 4-LUT Others 33.2% 45.5%

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
7

Automated
Processor MiBench
Transistor Sizing
HDL Description Application
Tool

M
COFEE 2.1

M
Cross-Compiled to the RISC-V ISA

M
Computational Unit GCC Cross-Compiler
BLE 1

M
D-FF

M
Capacitance Transistor Decomposition
D-FF

M
M
Resistance Sizing Xilinx ISE Component Utilization Extraction

M
M
M
D-FF

M
and Active and Leisure Periods

M
M
D-FF

M
M
Synthesis to RISC-V Full System Simulator (Spike)

M
*Area, Delay,

M
4-Input Functions

M
and Power Estimation
10 Hspice Berkeley ABC / Resource
(45nm PTM) Altera Quartus Utilization
M

I
M

BLE 5

M
Candidate
M

NPN-Class Extraction Switching

M
Architecture File Components Rate
D-FF
M

and Analysis

M
D-FF
M

Development
M

D-FF
M

Component
M

Mapping Algorithm
M

Delay
M

(Pre-Designed
M

Clustering and Low-Power CHLs)

Placement Application Run-Time
In-house C++ Estimation (Calculations)
VPR-7 Tool
M M M M M M M M M M M M M M M M In-house C++
M M M M M M M M M M M M M M M M BLE 8
M M M M M M M M M M M M M M M M
D-FF Power *Area and Area Application
Critical Path Calculations Power per RHL Calculations
D-FF Execution-Time
Delay
D-FF
Power Area

Fig. 11. Implementation and evaluation flow of the proposed architecture(s)

Fig. 10. Proposed Configurable Logic Block (CLB) TABLE VI
C ONFIGURABLE H ARD L OGICS AND LUT PARAMETERS

fragmentally-accessed ones. For this aim, each MiBench benchmark Area

Delay Static Power Config. Area
is cross-compiled to the RISC-V ISA using modified GCC [8]. Then, Cell (min width
(ps) (nW) Cell (µm2 )
transistors)
the utilization and access pattern to the functional units is obtained
CHL1 67 180 5 72 8.8
by the means of RISC-V ISA Simulator (Spike) [46]. Next, HDL
CHL2 69 180 5 72 8.8
description of each component is extracted from the processor HDL 4-LUT 155 1388 16 267 32.5
source code and then optimized and mapped to 4-input functions by
the means of Berkeley ABC synthesis tool [32] and Altera Quartus
[33]. The truth-table of each function is generated using an in- A. Preliminary
house C script and is fed to the Boolean Matcher [47] to extract
Table V demonstrates the architectural parameters and the descrip-
the corresponding NPN-Class representation. We exploit COFFE [34]
tion for the proposed and baseline architectures used in the VPR
for efficient transistor sizing and area and delay estimation of LUT-
experiments. Similar to commercial FPGA devices [49], [50], we
based architectures (according to commercial architectures), while
used the conventional island-style architecture exploited in VPR tool.
evaluation of the CHLs is performed using HSPICE circuit-level
To calculate the area, we first extracted total area from VPR post
simulations with Nangate 45nm Open Cell Library [48]. The area and
routing reports which uses minimum width transistor model [34]
delay values of each cell are fed to the VPR [35] architecture file for
represented by Equation 3 in which x is the ratio of transistor width
circuit placement on FPGA-based platform. To obtain more accurate
to minimum width. For instance, a minimum-sized inverter area in
and reliable results, we run VPR simulations using different initial
the 45nm technology can be obtained by setting x = 90nm 90nm
= 1 for
placements (i.e., different placement seeds) since the timing and area
the NMOS, and x = 135nm
90nm
for the PMOS transistor.
results are sensitive to initial placements. Lastly, area and critical √
path delay are extracted directly from VPR while power is estimated M in. width transistor(x) = 0.447 + 0.128x + 0.391 x (3)
by post-processing VPR-generated reports such as number of logic
While Equation 3 can be directly used for comparison purpose, to
blocks and routing switches (channel width). An in-house C script
obtain the geometric (i.e., layout) area, it is shown that multiplying
is used to estimate execution-time based on profiling the instruction
it to 60F 2 , in which F is the technology size, gives an acceptable
execution flow reported by RISC-V ISA Simulator (Spike). In the
layout-level area [34], [35].
C script, we have considered the switching time overhead which is
Area(x) = M in. width transistor(x) × 60F 2 (4)
needed whenever an RU should be reconfigured.
B. Area
TABLE V
PARAMETERS IN THE PROPOSED AND BASELINE RC ARCHITECTURES According to Table III, low utilized components are integrated into
3 RUs with 3152, 10918, and 2074 LUTs. Therefore, the number of
Parameter Definition Value
LUTs is decreased by 26.8% from 43743 (which is the total number
N Cluster (CLB) size 10 of LUTs in all functional units in RC/module) to 32019 (number
K LUT size 4
of LUTs in the proposed architecture). Taking into account that by
I Cluster inputs (from adjacent routing channels) 22
CB connectivity factor (determines no. of
reducing logic resources, routing resources are also reduced; hence,
Fcin 0.16 number of configuration bits in the proposed architecture is decreased
channel tracks connected to each CB)
Cluster output connectivity factor (determines no. by 30.9%, as shown in Fig. 12. Area of LUT-based architectures is
Fcout 0.1
of SBs a logic block output is connected to) first calculated with respect to the dimension-less area estimation
X LUT input multiplexer size 16 in term of minimum width transistor count, and then the layout-
Cluster feedback factor (determines no. of level area is obtained using Equation 4. The layout area of ASIC-
Ff b 5/16
other logic blocks output in LUT input mux) based structures is directly reported by Synopsys Design Compiler

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
8

in overall execution time of the application, LUT-RU architecture

outperforms the RC/module architecture by 6.3% improvement in
execution time, as shown in Fig. 13.

E. Power and Energy Consumption

The total static power for each architecture is obtained by post-
processing the VPR reports by summing the power of logic and
routing resources in the target architecture. Notice that we consider
the static power of both utilized and unutilized LUT/CHL cells in
power and energy computation despite that the RUs do not use all
cells. Fig. 14 compares the normalized static power consumption
Fig. 12. Comparing the number of configuration bits and area of RISC-V in different architectures. Compared with the baseline RC/module
components in the baseline and proposed architectures architecture, the proposed LUT-RU architecture reduces the power
consumption by 32.5%. The static power of both logic and routing
is also decreased since the LUT-RU architecture reduces both of
the logic and routing resources. In addition, while the proposed
architecture is most power-efficient in reconfigurable architectures
and reduces the number of logic resources, however, the conventional
ASIC implementation, i.e., ASIC/module, is still power-efficient than
all other architectures.
The energy consumption of different architectures are obtained by
Equation 5. Energy is estimated as summation of reconfiguration and
operation energy while reconfiguration energy is zero in ASIC-based
architectures. P owertotal stands for total power of RUs or total
power of resources in baseline architectures and non-reconfiguring
Fig. 13. Comparing the delay of functional units and execution time in components (i.e., Integer Unit). For reconfiguration, a write energy
different architectures in RISC-V processor of 3.5 × 10−18 J for a single SRAM cell in 45nm technology
is assumed [51]. Fig. 14 compares the normalized energy in the
and then is multiplied to the number of instances. According to Fig.
proposed and baseline architectures. As demonstrated in this figure,
12, the proposed LUT-RU architecture has reduced the area by 30.7%,
LUT-RU architecture improves the energy by 36.9% as compared to
compared with the RC/module baseline architecture. As expected, this
the baseline RC/module architecture.
improvement is lower than the improvement in the number of LUTs
∑N
since some CHL cells have been substituted with LUTs. E=( #SRAM (RUi ) × #Rcf g(RUi ) × ESRAM )
i=1
(5)
C. Critical Path Delay
+ P owertotal × Execution time
The delay of the ASIC-based architectures is obtained from Synop-
sys Design Compiler. As shown in Fig. 13, the LUT-RU architecture
does not affect the logic delay compared with the baseline RC/module
since once a component, e.g., FP-FP2F, is configured in either
RC/module or LUT-RU architectures, it would have identical delay
in both architectures due to the same mapping and logic resource
usages. Yet, the routing delay will be reduced since the channel
width in LUT-RU architecture is considerably less than RC/module.
Hence, the proposed LUT-RU architecture reduces critical path delay
by 9.2% compared to the baseline RC/module architecture.

D. Application Execution Time Fig. 14. Comparing the power and energy consumption of different architec-
To calculate the execution time, we multiplied execution time of tures implementing the RISC-V processor
each instruction to its corresponding component delay. In addition, It is noteworthy that the proposed architecture advances power-
for the proposed architectures, we added the reconfiguration time gating approach in different aspects. First, the most challenging issue
by calculating the write latency of the total number of configu- of power-gating is so-called inrush current, i.e., a substantial wake-
ration bits (sum of logic and routing) with the method proposed up current occurs whenever a power-gated module turns abruptly
in [18]2 . Accordingly, we assume a 128-bit parallel data bus [18] on, which may cause violation scenarios such as instability of the
and a write latency of 0.2ns for a single SRAM cell [51]. Thus, registers content and functional errors resulting from dynamic voltage
reconfiguration time of RU1, RU2, and RU3 in LUT-RU architecture (IR) drop [52]. It also can increase the wake-up time and wake-up
will be 424619×0.2ns
128
= 0.66 µs, 1481035×0.2ns
128
= 0.23 µs, and energy. Previous studies have reported inrush power contributes to
337531×0.2ns 20% of the static power in FPGAs with 45nm technology size [53].
128
= 0.52 µs, respectively, in which 424619, 1481035,
and 337531 are the configuration bit count of the largest component Second, power gating requires complicated controller and compiling
in each RU. Due to the fact that the configurations time is prorated techniques to (a) anticipate the idle intervals and evaluate whether
the power gating is advantageous in spite of associated overheads,
2 It is noteworthy configuring one component to another in [18] has been (b) predict data arrival in order to initiate the wake-up phase and
calculated based on configuring configuration cells of the smallest component assure that only a small fraction of resources is switched each time to
which is not correct. preclude timing (and hereby, functional) errors, and (c) route signals

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
9

TABLE VII
U TILIZATION , AREA , AND CHL S COVERAGE RATIO IN LEON2
TAP Peri Cache RAM 1 PROCESSOR FUNCTIONAL UNITS
Functional Area (4-LUT)
Utilization
Units RC/module LUT-RU CHL-LUT-RU
Mem ctrl RAM 2
Cache ADD 15.42% 176 176 4-LUT 176 4-LUT
MUL 3.4 × 10−4 % 485 414 CHL1, 61 CHL2
2498 4-LUT
DIV 9.49% 2498 , and 2023 4-LUT)
RAM 3
FPU
RAM 4
IU
DRAM 1 Adder Divider

DRAM 2 Multiplier

Low High Not

Memories
Utilization Utilization Considered

Fig. 15. LEON2 processor microarchitecture (the area of each block is Fig. 16. Utilization ratio of LEON2 components in MiBench benchmarks
normalized to the chip area.)

in such a way that they do not use the resources of power-gated

regions. Another advantage of the proposed architecture with respect
to power-gating is the improvement of reliability. When a soft-error
flips a configuration cell in the proposed architecture, this cell will
be automatically corrected in the next configurations (notice that the
power-gating methods do no turn-off the SRAM cells). Last but not
least, power-gating approach leaves the part of device unused, while
the proposed architecture utilizes all the resources which shortens the
wirelength and thereby, wire delays, consequently. Nevertheless, since Fig. 17. Average inter-component switching ratio in LEON2 processor
FPGAs comprise abundant routing resources which mainly remain
underutilized, power gating of the routing resources [54] can be
employed in tandem with our proposed architecture to further increase
the power efficiency.

F. Alternative Processor
We have examined the applicability of the proposed architecture
on an alternative open-source soft-core processor. We choose LEON2
which is SPARC-V8-compliant, as an alternative processor for several
reasons. First, it is popular among a wide range of embedded
processors and also compatible and extensible to be customized for
diverse range of performance/cost trade-off, e.g., LEON2-FT. Second,
the full HDL source code of total LEON2 (including FPU) is license Fig. 18. Comparing the number of LUTs and area of LEON2 components
free, while source code of FPU is not available in LEON3 and in the baseline and proposed architectures
LEON4 which causes problem in running application with floating
point instructions. Finally, in RISC-V case study, the integrated 16. Utilization ratio of the integer divider (DIV) is approximately
functional units all belong to FPU, but here in LEON2, as we will 3.48 × 10−6 % while it occupies a significant area (79.1%) of total
demonstrate in the following, all functional units are separate and functional units. Therefore, it can be efficiently integrated with integer
thereby, integration of low-utilized units is more challenging. multiplier (MUL) into a shared RU. Hence, switching ratio of the
Fig. 15 illustrates the synthesized floorplan of LEON2. As shown proposed RU unit is 2.81 × 10−6 on average, as demonstrated in
in this figure, LEON2 functional units include a 32-bit adder, 32- Fig. 17. Subsequently, reconfiguration overhead of the proposed RU
bit multiplier (that can be configured to 32x32, 32x16, 32x8, and will be small and prorated in application execution time, leaving
16x16), 32-bit divider, and an optional FPU. FPU as a co-processor negligible impact on total execution time of applications. The rest
is licensed and therefore, is not available in the processor source code. of the evaluation flow for the proposed and baseline architectures in
Therefore, we have omitted the FPU unit from our evaluations and LEON2 processor is the same as RISC-V processor.
focused on the integer functional units which proves the generality of 1) Area: Fig. 18 demonstrates the number of LUTs and configu-
the proposed architectures not just in FPU. Accordingly, in the rest ration bits in the proposed and baseline architectures. As shown in
of this section, the reported results for LEON2 processor are in the this figures, the number of LUTs and configuration bits in LUT-RU
scope of Integer Unit (IU). has been reduced by 23.1%, compared to the baseline RC/module
To identify the utilization and access pattern to the functional architecture. The total area for LUT-RU is also reduced by 23.1%.
units, each MiBench benchmark is cross-compiled to the SPARC- The detailed results for each architecture are shown in Fig. 18.
V8 architecture using RTEMS [55]. Afterwards, Simics full system 2) Power and Energy Consumption: As demonstrated in Fig. 19,
simulator [56] is used to obtain the utilization and access patterns static power of the LUT-RU architecture is reduced by 23.2% as
of the functional units which are shown in Table VII and Fig. compared to the RC/module architecture. In addition, compared to

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
10

TABLE VIII
R ESULTS FOR CHL-LUT-RU IN RISC-V AND LEON2 PROCESSORS
NORMALIZED TO RC/ MODULE ( CONSIDERING LOGIC RESOURCES )

Parameters/ RISC-V LEON2

Achitectures LUT-RU CHL-LUT-RU LUT-RU CHL-LUT-RU
# of Config. Bits 0.72 0.65 0.78 0.69
Area 0.72 0.59 0.78 0.66
Critical Path 1.04 0.93 0.93 0.89
Execution Time 1.06 0.97 0.94 0.89
Energy 0.71 0.48 0.78 0.55
Static Power 0.75 0.47 0.72 0.53

of target CHL or LUT. Static power of each cell is measured using

Fig. 19. Comparing the power and energy of different architectures imple- HSPICE simulations with the same transistor technology files and
menting LEON2 processor
temperature used in Nangate library (i.e., 25 °C).
The summary of results is reported in Table VIII. As demonstrated
in this table, the proposed CHL-LUT-RU architecture has further
reduced the number of SRAM configuration bits, area, static power,
critical path delay, execution time, and energy compared to the
proposed LUT-RU in the scope of logic resources. Note that reducing
configuration bits can be translated as the circuit reliability due to the
reduction of the number of susceptible bits to soft errors [57], [58].
In addition, while our proposed architectures, particularly CHL-LUT-
RU, have significantly reduced the logic resources, the conventional
ASIC implementation, i.e., ASIC/module, is still power-efficient than
all other architectures. This is mainly due to the large number of
Fig. 20. Comparing the delay of functional units and execution time in
different architectures implementing LEON2 processor configuration cells used in the other architectures while the ASIC
implementation of a 4-input Boolean function consumes only 35nW
the RC/module architecture, LUT-RU architecture has reduced energy on average, whereas this value is greater than 180nW for CHL cells.
by 23.2%. These results are expectable since Divider is considerably
smaller than Multiplier. Hence, the improvements are nearly equal VI. D ISCUSSION
between energy-power and between area-configuration bits. Here, we discuss some points related to the proposed architecture.
3) Critical Path Delay and Application Execution Time: Fig. 20 1) Routing architecture: We have not investigated the routing
indicates the critical path delay and total execution time of the architecture for CHL-LUT-RU due to its high complexity in both
baseline and proposed architectures. The results demonstrate that total implementation and analysis. However, since the routing resources
execution time has only increased by 10−5 % in LUT-RU compared in a typical island style FPGA architecture are surrounded by logic
with the RC/module architecture. This is due to negligible utilization clusters, reducing the number of required logic clusters by the factor
ratio of the divider which reduces the reconfiguration time overhead. of S will reduce the number of routing resources (i.e., switch boxes)
It should be noted that the efficiency of the proposed architecture and thereby their area and power by the same factor. In addition,
is irrelevant to the size of processor itself, as we have shown its taking the routing delay into account, the critical path delay will not
effectiveness in the moderate-size RISC-V processor with ∼ 35, 000 be increased since at each time only one of the functional units is
LUTs, and a smaller one (LEON2) consists of ∼ 3, 200 LUTs. mapped on the corresponding RU owning the same (or more in the
The key prerequisite for effectiveness of the proposed architecture case that is smaller than RU) routing and logic resources as in its
is (a) low utilized units with fragmented access pattern, and (b) low baseline architecture. In particular, we expect that the critical path
utilized units have close area. Nevertheless, larger processors result delay may also be improved since higher number of resources (more
in larger RUs and consequently, higher number of reconfiguration flexibility to place and route) will be allocated to specific components.
bits and therefore, reconfiguration time. The proposed architecture Furthermore, in the proposed architecture, proximity of the logic
also improves the critical path delay which will directly enhance the clusters of each component may increase after the mapping due to
execution time due to the shorter clock cycles which is valid in RISC- the exclusive mapping of components and reduction of the array size.
V case study. It can further improve the routing delay.
2) Optimal heterogeneous mapper: A heterogeneous architecture
G. Evaluating CHL-LUT-RU such as the proposed CHL-LUT-RU may need modifications in the
In the proposed CHL-LUT-RU architecture, each shared LUT that original mapping algorithm that is used in a homogeneous, pure 4-
its function can be implemented with either CHL1 or CHL2, is simply LUT based architecture [43]; therefore, detailed investigation of the
substituted by an appropriate cell in the netlist BLIF file. The CHL- intra-cluster architecture (e.g., ratio of CHL to LUTs) and mapping
LUT-RU architecture has been evaluated after placement stage and the to such an architecture is required. We will cover details of both the
results are reported in terms of logic resources. Since the contribution routing and logic architectures as an extension of this work.
of this work is focused on logic architecture, the impact of routing 3) Reconfiguration overhead: Integrating underutilized compo-
resources has not been considered. Nevertheless, a brief discussion nents into shared RUs can impose reconfiguration overhead which
about routing is represented in Section VI. We intend to consider can be problematic in hard real-time designs. Although the impact of
proposing a customized (generic) routing architecture in the future reconfiguration time on the total application runtime is negligible
track of this work. Table VI reports area, delay, and static power of (about 2%), it can be eliminated using the concept of shadow
4-LUT and CHLs which are obtained using HSPICE simulations and SRAM cells [59]. Nevertheless, a trade-off between eliminating the
with transistor sizing and sub-circuit models obtained from COFFE. reconfiguration delay and area-power overhead of shadow SRAMs
The delays reported in Table VI are the average delay of all inputs should be taken into account.

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
11

TABLE IX
S UMMARY OF THE RESULTS FOR VARIOUS SYNTHESIS TOOLS AND PROCESSORS ACROSS ALL BENCHMARK ( NORMALIZED TO RC/ MODULE ).
Parameter Area Critical Path Execution Time Energy Static Power
Processor RISC-V LEON2 RISC-V LEON2 RISC-V LEON2 RISC-V LEON2 RISC-V LEON2
Berkeley Altera Berkeley Berkeley Altera Berkeley Berkeley Altera Berkeley Berkeley Altera Berkeley Berkeley Altera Berkeley
Synthesis Tool
ABC Quartus ABC ABC Quartus ABC ABC Quartus ABC ABC Quartus ABC ABC Quartus ABC
ASIC/inst
0.06 0.06 0.05 0.16 0.16 0.12 0.01 0.01 0.26 0.01 4.18 × 10−3 0.3 0.41 0.26 0.25
(logic)
ASIC/module −3
0.01 0.01 0.01 1.8 × 10 0.16 0.57 0.17 0.16 1.18 0. 01 0.01 0.11 0.07 0.07 0.09
(logic)
RC/all 0.29 0.24 0.64 0.93 1.02 0.99 41.48 29.57 46.81 0.1 0.08 0.64 4.2 × 10−3 2.7 × 10−3 0.03
RC/module 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
LUT-RU 0.69 0.71 0.77 0.91 0.97 1 0.94 0.99 1 0.63 0.69 0.77 0.68 0.69 0.77

4) Applicability with Non-Volatile Memories (NVM): Although VII. C ONCLUSION

our proposed architecture is targeted for SRAM-based soft-core pro-
In this paper, we presented a reconfigurable architecture for em-
cessors, it is definitely not limited to SRAMs and can be expanded to
bedded soft-core processors, which aims to improve area and power
NVM platforms as well. However, despite promising characteristics
efficiency of FPGA-based soft-core processors while maintaining
of NVM such as negligible leakage power, they encounter with
flexibility to customize and update the processor for designers. In the
several unaddressed deficiencies such as write latency and energy and
proposed architecture, we integrated low-utilized and fragmentally-
still are in prototyping phase. In addition, the proposed architecture
accessed components into a single RU and used efficient configurable
may require large number of reconfigurations which is not consistent
logics along with the mapping algorithm to implement the overlap-
with currently-available NVMs.
ping functions of functional components mapped to each RU. Our
5) Applicability in emerging applications: One of the promising integration is based on carefully profiling the frequency and access
application of the proposed architecture can be in implementing Ma- pattern of instructions in MiBench applications, in order to well select
chine Learning (ML) algorithms, e.g., Neural Networks. Despite the the merged components. We have implemented RISC-V processor to
widespread research on efficient implementation of Neural Networks, examine the efficiency of the proposed architectures. Experimental
only a limited range of objective functions, especially sum or mean results show that, as compared to conventional FPGA-based soft-core
squared error used in linear regression, has been implemented on processors, the proposed architecture improve the area footprint, static
the state-of-the-art FPGAs [60]. This is because, operations such power, energy consumption, critical path delay, and total execution
as division and square root used in the gradient descent cycle of time by 30.7%, 32.5%, 36.9%, 9.2%, and 6.3%, respectively.
objective function either occur infrequently or consume significant
amount of resources. The proposed architecture can be beneficial R EFERENCES
in these applications when for example, floating-point multiplier
can be merged with divider since in ML algorithms, a stream of [1] D. D. Gajski and F. Vahid, “Specification and design of embedded
multiplication-additional (MAC) operations is followed by a division. hardware-software systems,” Design & Test of Computers, IEEE, vol. 12,
Therefore, the context-switching inside an RU will be tolerable. As an no. 1, pp. 53–67, 1995.
[2] P. Marwedel, Embedded system design: Embedded systems foundations
interesting point, in the proposed architecture we have already merged of cyber-physical systems. Springer Science & Business Media, 2010.
these units within RISC-V processor. The proposed architecture is [3] M. Barr, Programming embedded systems in C and C++. ” O’Reilly
also applicable for component-based designs other than soft-core pro- Media, Inc.”, 1999.
cessors, such as DSP cores that are exploited in ML algorithms [60]. [4] P. Marwedel and G. Goossens, Code generation for embedded proces-
sors. Springer Science & Business Media, 2013, vol. 317.
We have examined the applicability of the proposed architecture on a
[5] I. Kuon and J. Rose, “Measuring the gap between FPGAs and ASICs,”
floating point DSP core from OpenCores [61] which its contributing Computer-Aided Design of Integrated Circuits and Systems, IEEE Trans-
components are listed in Table X. Depending on the distribution of actions on, vol. 26, no. 2, pp. 203–215, 2007.
the operations (tunable according to the target application), several [6] M. Sarrafzadeh, F. Dabiri, R. Jafari, T. Massey, and A. Nahapetan, “Low
components of the DSP core [60] can be merged into a shared RU. power light-weight embedded systems,” in Low Power Electronics and
Design. ISLPED’06. Proceedings of the 2006 International Symposium
One appropriate integration to save considerable area and power can on. IEEE, 2006, pp. 207–212.
be achieved by merging the divider, multiplier, and subtractor units. [7] (Accessed March, 2016) Gaisler research website. [Online]. Available:
Our experimental results for this DSP core reveal that area is saved www.gaisler.com/
by 50.2%, with 50.1% improvement in static power while keeping [8] K. Asanović and D. A. Patterson, “Instruction sets should be free:
the latency intact. The case for RISC-V,” Technical report, University of California at
Berkeley, http://www. eecs. berkeley. edu/Pubs/TechRpts/2014/EECS-
Another use-case of the proposed method can be in FPGA-assisted 2014-146. pdf, Tech. Rep., 2014.
Cloud computing, a.k.a, FPGA virtualization [62] which incorporates [9] (Accessed March, 2016) Opensparc website. [Online]. Available:
FPGAs in datacenters in order to improve the processing capacity www.opensparc.org
and power consumption. A challenging task in such platforms is [10] (Accessed March, 2016) Openrisc website. [Online]. Available:
an efficient use of FPGA resources since an FPGA is expected www.openrisc.io
[11] F. Plavec, B. Fort, Z. G. Vranesic, and S. D. Brown, “Experiences with
to afford computation of different users (i.e., several tasks) at the soft-core processor design,” in null. IEEE, 2005, p. 167b.
same time. Since the proposed architecture integrates low-utilized [12] L. Barthe, L. V. Cargnini, P. Benoit, and L. Torres, “The SecretBlaze: A
functional units into a single RU, provided that we keep the FPGA configurable and cost-effective open-source soft-core processor,” in Par-
size the same, further tasks can be implemented into the same FPGA allel and Distributed Processing Workshops and Phd Forum (IPDPSW),
International Symposium on. IEEE, 2011, pp. 310–313.
device. This will reduce the cost-per-user or cost-per-task.
[13] J. G. Tong, I. D. Anderson, and M. A. Khalid, “Soft-core processors
TABLE X for embedded systems,” in Microelectronics. ICM’06. International
COMPONENTS AREA IN THE DSP CORE [61] Conference on. IEEE, 2006, pp. 170–173.
[14] P. Yiannacouras, J. Rose, and J. G. Steffan, “The microarchitecture of
Functional Unit Adder Subtractor Round Divider Multiplier Exception
# of 4-LUTs 689 2667 434 3723 2688 1067
fpga-based soft processors,” in Proceedings of the 2005 international
conference on Compilers, architectures and synthesis for embedded
systems. ACM, 2005, pp. 202–212.

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
12

[15] J. Bachrach, H. Vo, B. Richards, Y. Lee, A. Waterman, R. Avižienis, [36] M. Hosseinabady and J. L. Nunez-Yanez, “Run-time power gating in hy-
J. Wawrzynek, and K. Asanović, “Chisel: constructing hardware in a brid arm-fpga devices,” in Field Programmable Logic and Applications
scala embedded language,” in Proceedings of the 49th Annual Design (FPL), 24th International Conference on. IEEE, 2014, pp. 1–6.
Automation Conference. ACM, 2012, pp. 1216–1225. [37] J. Teich, “Hardware/software codesign: The past, the present, and
[16] A. A. Bsoul and S. J. Wilton, “An FPGA architecture supporting dy- predicting the future,” Proceedings of the IEEE, vol. 100, no. Special
namically controlled power gating,” in Field-Programmable Technology Centennial Issue, pp. 1411–1430, 2012.
(FPT), 2010 International Conference on. IEEE, 2010, pp. 1–8. [38] VectorBlox/risc-v. VectorBlox Computing Inc. [Online]. Available:
[17] H. Esmaeilzadeh, E. Blem, R. St Amant, K. Sankaralingam, and https://github.com/VectorBlox/orca
D. Burger, “Dark silicon and the end of multicore scaling,” in ACM [39] Y. Lee, A. Waterman, R. Avizienis, H. Cook, C. Sun, V. Stojanović,
SIGARCH Computer Architecture News, vol. 39, no. 3. ACM, 2011, and K. Asanović, “A 45nm 1.3 ghz 16.7 double-precision gflops/w risc-
pp. 365–376. v processor with vector accelerators,” in European Solid State Circuits
[18] A. R. Ashammagari, H. Mahmoodi, and H. Homayoun, “Exploiting Conference (ESSCIRC), 2014-40th. IEEE, 2014, pp. 199–202.
STT-NV technology for reconfigurable, high performance, low power, [40] N. Muralimanohar, R. Balasubramonian, and N. P. Jouppi, “Cacti 6.0:
and low temperature functional unit design,” in Proceedings of the A tool to model large caches,” HP Laboratories, pp. 22–31, 2009.
conference on Design, Automation & Test in Europe. European Design [41] H. Wong, V. Betz, and J. Rose, “Comparing FPGA vs. custom CMOS
and Automation Association, 2014, p. 335. and the impact on processor microarchitecture,” in Proceedings of the
[19] A. R. Ashammagari, H. Mahmoodi, T. Mohsenin, and H. Homayoun, 19th ACM/SIGDA international symposium on Field programmable gate
“Reconfigurable STT-NV LUT-based functional units to improve perfor- arrays. ACM, 2011, pp. 5–14.
mance in general-purpose processors,” in Proceedings of the 24th edition [42] E. Ahmed and J. Rose, “The effect of lut and cluster size on deep-
of the great lakes symposium on VLSI. ACM, 2014, pp. 249–254. submicron fpga performance and density,” IEEE Transactions on Very
[20] H. Singh, M.-H. Lee, G. Lu, F. J. Kurdahi, N. Bagherzadeh, and Large Scale Integration (VLSI) Systems, vol. 12, no. 3, pp. 288–298,
M. C. Eliseu Filho, “Morphosys: an integrated reconfigurable system for 2004.
data-parallel and computation-intensive applications,” Computers, IEEE [43] I. Ahmadpour, B. Khaleghi, and H. Asadi, “An efficient reconfigurable
Transactions on, vol. 49, no. 5, pp. 465–481, 2000. architecture by characterizing most frequent logic functions,” in Field
[21] B. Mei, S. Vernalde, D. Verkest, H. De Man, and R. Lauwereins, Programmable Logic and Applications (FPL), 2015 25th International
“ADRES: An architecture with tightly coupled VLIW processor and Conference on. IEEE, 2015, pp. 1–6.
coarse-grained reconfigurable matrix,” in Field Programmable Logic and [44] A. Ahari, B. Khaleghi, Z. Ebrahimi, H. Asadi, and M. B. Tahoori,
Application. Springer, 2003, pp. 61–70. “Towards dark silicon era in fpgas using complementary hard logic
[22] S. C. Goldstein, H. Schmit, M. Budiu, S. Cadambi, M. Moe, and design,” in Field Programmable Logic and Applications (FPL), 2014
R. R. Taylor, “Piperench: A reconfigurable architecture and compiler,” 24th International Conference on. IEEE, 2014, pp. 1–6.
Computer, vol. 33, no. 4, pp. 70–77, 2000. [45] Z. Ebrahimi, B. Khaleghi, and H. Asadi, “PEAF: A Power-Efficient
[23] R. Lysecky, G. Stitt, and F. Vahid, “Warp processors,” in ACM Transac- Architecture for SRAM-Based FPGAs Using Reconfigurable Hard Logic
tions on Design Automation of Electronic Systems (TODAES), vol. 11, Design in Dark Silicon Era,” Computers, IEEE Transactions on, In press,
no. 3. ACM, 2004, pp. 659–681. 2017.
[46] (Accessed August, 2016) Behavioural simulation (spike). [Online].
[24] Z. A. Ye, A. Moshovos, S. Hauck, and P. Banerjee, CHIMAERA:
Available: www.lowrisc.org/docs/untether-v0.2/spike
a high-performance architecture with a tightly-coupled reconfigurable
[47] D. Chai and A. Kuehlmann, “Building a better boolean matcher and sym-
functional unit. ACM, 2000, vol. 28, no. 2.
metry detector,” in Proceedings of the conference on Design, automation
[25] J. E. Carrillo and P. Chow, “The effect of reconfigurable units in
and test in Europe: Proceedings. European Design and Automation
superscalar processors,” in Proceedings of the 2001 ACM/SIGDA ninth
Association, 2006, pp. 1079–1084.
international symposium on Field programmable gate arrays. ACM,
[48] (Accessed March, 2016) Nangate Open Cell Library. [Online].
2001, pp. 141–150.
Available: www.nangate.com/
[26] M. Labrecque and J. G. Steffan, “Improving pipelined soft processors
[49] “Stratix-2 platform FPGA hand book,” Altera, April 2011.
with multithreading,” in Field Programmable Logic and Applications,
[50] “Virtex-4 platform FPGA user guide,” Xilinx, December 2008.
FPL. International Conference on. IEEE, 2007, pp. 210–215.
[51] A. Chen, J. Hutchby, V. Zhirnov, and G. Bourianoff, Emerging nano-
[27] Z. Hu, A. Buyuktosunoglu, V. Srinivasan, V. Zyuban, H. Jacobson, and electronic devices. John Wiley & Sons, 2014.
P. Bose, “Microarchitectural techniques for power gating of execution [52] A. A. Bsoul and S. J. Wilton, “A configurable architecture to limit
units,” in Proceedings of the 2004 international symposium on Low wakeup current in dynamically-controlled power-gated fpgas,” in Pro-
power electronics and design. ACM, 2004, pp. 32–37. ceedings of the ACM/SIGDA international symposium on Field Pro-
[28] S. Roy, N. Ranganathan, and S. Katkoori, “A framework for power- grammable Gate Arrays. ACM, 2012, pp. 245–254.
gating functional units in embedded microprocessors,” IEEE transactions [53] S. Sharp. Power management solution guide - xilinx. [On-
on very large scale integration (VLSI) systems, vol. 17, no. 11, pp. 1640– line]. Available: https://www.xilinx.com/publications/archives/solution
1649, 2009. guides/power management.pdf
[29] A. A. Bsoul, S. J. Wilton, K. H. Tsoi, and W. Luk, “An fpga architecture [54] Z. Seifoori, B. Khaleghi, and H. Asadi, “A power gating switch box
and cad flow supporting dynamically controlled power gating,” IEEE architecture in routing network of sram-based fpgas in dark silicon
Transactions on Very Large Scale Integration (VLSI) Systems, vol. 24, era,” in Proceedings of the Conference on Design, Automation & Test
no. 1, pp. 178–191, 2016. in Europe. European Design and Automation Association, 2017, pp.
[30] P. Yiannacouras, J. G. Steffan, and J. Rose, “Exploration and cus- 1342–1347.
tomization of FPGA-based soft processors,” Computer-Aided Design of [55] (Accessed August, 2016) RTEMS Cross Compilation System (RCC).
Integrated Circuits and Systems, IEEE Transactions on, vol. 26, no. 2, [Online]. Available: www.rtems.org
pp. 266–277, 2007. [56] P. S. Magnusson, M. Christensson, J. Eskilson, D. Forsgren, G. Hallberg,
[31] M. R. Guthaus, J. S. Ringenberg, D. Ernst, T. M. Austin, T. Mudge, and J. Hogberg, F. Larsson, A. Moestedt, and B. Werner, “Simics: A full
R. B. Brown, “MiBench: A free, commercially representative embedded system simulation platform,” Computer, vol. 35, no. 2, pp. 50–58, 2002.
benchmark suite,” in Workload Characterization, 2001. WWC-4. IEEE [57] S. Yazdanshenas, H. Asadi, and B. Khaleghi, “A scalable dependability
International Workshop on. IEEE, 2001, pp. 3–14. scheme for routing fabric of sram-based reconfigurable devices,” IEEE
[32] BLSG, “ABC: A system for sequential synthesis and verification,” Trans. VLSI Syst., vol. 23, no. 9, pp. 1868–1878, 2015.
Berkeley Logic Synthesis and Verification Group, 2011. [58] H. Asadi and M. B. Tahoori, “Analytical techniques for soft error rate
[33] J. Pistorius, M. Hutton, A. Mishchenko, and R. Brayton, “Benchmarking modeling and mitigation of fpga-based designs,” IEEE Trans. VLSI Syst.,
method and designs targeting logic synthesis for fpgas,” in Proc. IWLS, vol. 15, no. 12, pp. 1320–1331, 2007.
vol. 7, 2007, pp. 230–237. [59] W. Zhang, N. K. Jha, and L. Shang, “Low-power 3D nano/CMOS hybrid
[34] S. Yazdanshenas and V. Betz, “Automatic circuit design and modelling dynamically reconfigurable architecture,” ACM Journal on Emerging
for heterogeneous fpgas,” in Field-Programmable Technology (FPT), Technologies in Computing Systems (JETC), vol. 6, no. 3, p. 10, 2010.
2017 International Conference on. IEEE, 2017, pp. 9–16. [60] D. Mahajan, J. Park, E. Amaro, H. Sharma, A. Yazdanbakhsh, J. K.
[35] J. Luu, J. Goeders, M. Wainberg, A. Somerville, T. Yu, K. Nasartschuk, Kim, and H. Esmaeilzadeh, “Tabla: A unified template-based framework
M. Nasr, S. Wang, Liu et al., “VTR 7.0: Next generation architecture for accelerating statistical machine learning,” in High Performance
and CAD system for FPGAs,” ACM Transactions on Reconfigurable Computer Architecture (HPCA), 2016 IEEE International Symposium
Technology and Systems (TRETS), vol. 7, no. 2, p. 6, 2014. on. IEEE, 2016, pp. 14–26.

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCAD.2018.2812118, IEEE
Transactions on Computer-Aided Design of Integrated Circuits and Systems
13

[61] (Accessed November, 2017) Opencores. [Online]. Available: www. Hossein Asadi (M’08, SM’14) received the B.Sc.
opencores.org and M.Sc. degrees in computer engineering from the
[62] S. Yazdanshenas and V. Betz, “Quantifying and mitigating the costs SUT, Tehran, Iran, in 2000 and 2002, respectively,
of fpga virtualization,” in Field Programmable Logic and Applications and the Ph.D. degree in electrical and computer
(FPL), 27th International Conference on. IEEE, 2017, pp. 1–7. engineering from Northeastern University, Boston,
MA, USA, in 2007.
He was with EMC Corporation, Hopkinton, MA,
USA, as a Research Scientist and Senior Hardware
Engineer, from 2006 to 2009. From 2002 to 2003, he
was a member of the Dependable Systems Labora-
tory, SUT, where he researched hardware verification
techniques. From 2001 to 2002, he was a member of the Sharif Rescue
Robots Group. He has been with the Department of Computer Engineering,
SUT, since 2009, where he is currently a tenured Associate Professor. He
is the Founder and Director of the Data Storage, Networks, and Processing
(DSN) Laboratory, Director of Sharif High-Performance Computing (HPC)
Center, the Director of Sharif Information and Coummnications Technology
Center (ICTC), and the President of Sharif ICT Innovation Center. He spent
three months in the summer 2015 as a Visiting Professor at the School of
Sajjad Tamimi has received the B.Sc. and M.Sc.
Computer and Communication Sciences at the Ecole Poly-technique Federele
degrees in computer engineering from Iran Univer-
de Lausanne (EPFL). He is also the co-founder of HPDS corp., designing
sity of Science and Technology (IUST) and Sharif
and fabricating midrange and high-end data storage systems. He has authored
University of Technology (SUT), Tehran, Iran, in
and co-authored more than eighty technical papers in reputed journals and
2014 and 2016, respectively. He has been with
conference proceedings. His current research interests include data storage
the Data Storage, Processing, and Networks (DSN)
systems and networks, solid-state drives, operating system support for I/O
Laboratory at the Department of Computer Engi-
and memory management, and reconfigurable and dependable computing.
neering, SUT, as a research assistant for two years.
Dr. Asadi was a recipient of the Technical Award for the Best Robot
His current research interests include reconfigurable
Design from the International RoboCup Rescue Competition, organized by
computing, embedded system design, and computer
AAAI and RoboCup, a recipient of Best Paper Award at the 15th CSI
architecture.
Internation Symposium on Computer Architecture & Digital Systems (CADS),
the Distinguished Lecturer Award from SUT in 2010, the Distinguished
Researcher Award and the Distinguished Research Institute Award from SUT
in 2016, and the Distinguished Technology Award from SUT in 2017. He is
also recipient of Extraordinary Ability in Science visa from US Citizenship
and Immigration Services in 2008. He has also served as the publication
chair of several national and international conferences including CNDS2013,
AISP2013, and CSSE2013 during the past four years. Most recently, he has
served as a Guest Editor of IEEE Transactions on Computers, an Associate
Editor of Microelectronics Reliability, a Program Co-Chair of CADS2015,
and the Program Chair of CSI National Computer Conference (CSICC2017).

Zahra Ebrahimi has received the B.Sc. and M.Sc.

degrees in computer engineering SUT, Tehran, Iran,
in 2014 and 2016, respectively. She has been with
the DSN Laboratory at the Department of Computer
Engineering, SUT, as a research assistant for four
years. Her current research interests include recon-
figurable computing and computer-aided design.

Behnam Khaleghi has received his B.Sc. and M.Sc.

degrees in computer engineering from SUT, Tehran,
Iran, in 2013 and 2016, respectively. He is currently
working as a research assistant in the DSN Labo-
ratory at the Department of Computer Engineering,
SUT. He spent the summer 2014 and 2015 as a re-
search assistant at the Chair for Embedded Systems
in the Karlsruhe Institute of Technology (KIT). His
research interests include reconfigurable computing,
CAD, and reliable system design. He has two Best
Paper Nominations at the DAC’17 and DATE’17.

0278-0070 (c) 2018 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.