Virtual Ways: Efficient Coherence
for Architecturally Visible Storage
in Automatic Instruction Set Extensions
Theo Kluter1,5, Samuel Burri2, Philip Brisk4,
Edoardo Charbon2,3, and Paolo Ienne1
1Ecole Polytechnique F´ ed´ erale de Lausanne (EPFL), School of Computer and
Communication Sciences, CH-1015 Lausanne, Switzerland
2Ecole Polytechnique F´ ed´ erale de Lausanne (EPFL), School of Engineering,
CH-1015 Lausanne, Switzerland
3Delft University of Technology, Circuits and Systems Group,
NL-2600 AA Delft, The Netherlands
4University of California, Riverside, Department of Computer Science and
Engineering, Riverside, CA 92521, USA
5Bern University of Applied Sciences, EKT, Microlab, Quellgasse 21,
CH-2501 Biel/Bienne, Switzerland
Abstract. Customizable processors augmented with application-specif-
ic Instruction Set Extensions (ISEs) have begun to gain traction in re-
cent years. The most effective ISEs include Architecturally Visible Storage
(AVS), compiler-controlled memories accessible exclusively to the ISEs.
Unfortunately, the usage of AVS memories creates a coherence prob-
lem with the data cache. A multiprocessor coherence protocol can solve
the problem, however, this is an expensive solution when applied in a
uniprocessor context. Instead, we can solve the problem by modifying
the cache controller so that the AVS memories function as extra ways of
the cache with respect to coherence, but are not generally accessible as
extra ways for use under normal software execution. This solution, which
we call Virtual Ways is less costly than a hardware coherence protocol,
and eliminate coherence messages from the system bus, which improves
energy consumption. Moreover, eliminating these messages makes Vir-
tual Ways significantly more robust to performance degradation when
there is a significant disparity in clock frequency between the processor
and main memory.
tion Set Extensions, Virtual Ways.
Extensible processors are a cost-effective platform that can help embedded sys-
tem designers meet their targets for performance and energy efficiency. These
Y.N. Patt et al. (Eds.): HiPEAC 2010, LNCS 5952, pp. 126–140, 2010.
c ? Springer-Verlag Berlin Heidelberg 2010
Virtual Ways: Efficient Coherence for Architecturally Visible Storage127
processors are augmented with application-specific custom instruction set exten-
sions (ISEs) that improve performance and energy efficiency for critical loops in
embedded applications. ISEs can be identified automatically [11,4], and a system
designer must only verify the ISEs and their interface to the processor, as the pro-
cessor itself has been pre-verified by the vendor. Although extensible processors
cannot compete with application-specific integrated circuits (ASICs) in terms of
performance and energy efficiency, they offer an economic advantage through a
simplified design and verification process and a reduced time-to-market.
To increase performance, ISEs have been augmented with Architecturally Vis-
ible Storage (AVS), which can be registers or compiler-controlled memories .
AVS memories are distinct from the cache hierarchy, and Directed Memory Ac-
cess (DMA) transfers move data between main memory and the AVS, bypassing
the caches, which creates a coherence problem. Kluter et al.  solved the coher-
ence problem using a snoopy hardware coherence protocol, which was designed
for use in multiprocessor systems. This solution has two drawbacks: area over-
head, and performance degradation due to coherence messages on the system
bus competing with off-chip memory accesses.
Virtual Ways, presented here, is a scheme by which the cache controller is
modified to ensure coherence between the data cache and the AVS memory. Un-
der this scheme, the data cache and AVS memory share a common interface,
and prefetch instructions are used in lieu of DMA transfers. A relaxed form of
inclusion between the data cache and AVS memory provides coherence: the data
in the AVS memory is always a subset of the data in the cache, but writes in
the AVS memory are not automatically written through to the cache, but writes
to the AVS memory (cache) and not written through to the cache (AVS mem-
ory). The cache controller, therefore, evolves into a low-cost hardware coherence
protocol for this specific case.
Virtual Ways and Speculative DMA are compared using a standard cell de-
sign flow to estimate the area overhead of the memory subsystem. For JPEG
compression, in which the AVS memory is a 64-entry register file containing
8-bit registers, and 8 read and 8 write ports, the area overhead of Speculative
DMA was 1.29x, due to the cost of the coherence protocol, including the AVS
memory, while the area overhead of Virtual Ways was 1.09x, due mostly to the
AVS memory and a slightly larger data cache state machine.
Virtual Ways and Speculative DMA are compared using an FPGA-based soft
processor emulation system to measure task latency and memory system energy
consumption. The experiments include a detailed case study of JPEG compres-
sion, and an evaluation of four EEMBC consumer V2 benchmarks: CJPEGV2
(compression), MPEG encoding and decoding, and AES. The most significant
result is that the speedups achieved by Speculative DMA degrade significantly as
the frequency of the processor increases while the frequency of off-chip memory
remains constant, whereas, Virtual Ways does not suffer from any noticeable per-
formance degradation. For CJPEGV2 and MPEG encoding and decoding, Vir-
tual Ways achieved a higher speedup and reduced energy consumption compared
128T. Kluter et al.
to Speculative DMA, while the results for both metrics were equal for AES, due
to the fact that all data structures in the AVS memories are read-only.
We performed a detailed analysis and case study of an internally-modified
version of JPEG compression that only compresses one color component; we
call this version ”JPEG” for simplicity. The analysis of JPEG includes a kernel-
by-kernel breakdown of the task latency and energy consumption of the two
techniques, and include a design space exploration in which the size and as-
sociativity of the instruction and data caches are varied. In the former study,
Virtual Ways achieves a significant energy reduction in the Quantisation kernel,
while achieving comparable task latencies across all kernels. In the latter study,
Virtual Ways reduces task latency and energy consumption, compared to Specu-
lative DMA, for each configuration. Looking across configurations, Virtual Ways
generally achieves the best results; however, a handful of the best performing
configurations of Speculative DMA do achieve better task latency and/or energy
consumption than the worst performing configurations of Speculative DMA; the
overall trend, however, favors Virtual Ways.
The remainder of the paper is organized as follows: Section 2 details related
work in the domain. Section 3 introduces Virtual Ways and describes their im-
plementation in an extensible processor featuring AVS-enhanced ISEs. Section 4
describes an FPGA-based soft processor emulation system that we use for our
performance evaluation, and Section 5 presents an in-depth case study using
JPEG compression, followed by a more general study using EEMBC consumer
V2 benchmarks. Section 6 concludes the paper.
2 Related Work
In early work on ISEs, the processor’s register file was the I/O interface [4,11].
A typical register file has two read ports and one write port, which limit the
size of each ISE and the attainable speedup. Multi-cycle ISEs [1,10,12,14,15]
overlap computation with I/O operations; however, the I/O interface remained a
bottleneck. Several microarchitectural modifications have successfully improved
input bandwidth, including shadow registers , register file clustering , and
utilizing the pipeline forwarding logic . Although generally effective, these
techniques do not improve output bandwidth, and data is transferred to the ISE
logic on the granularity of scalar variables; they do not support bulk transfers
Biswas et al.  introduced architecturally visible storage (AVS), which was
limited to small ROMs that hold constant values, and state registers. In a subse-
quent work, they augmented their ISEs with small compiler-controlled memories
that hold arrays. DMA transfers move data into and out of the AVS memories
bypassing the cache . This solution is similar to scratchpad memories
which are also placed under compiler control. Scratchpad memories have been
proposed as an alternative to caches for embedded systems, because eliminating
the tag array reduces per-access energy consumption, and deterministic hit/miss
behavior improves predictability of worst-case execution time.
Virtual Ways: Efficient Coherence for Architecturally Visible Storage129
AVS memories, in contrast, co-exist with caches, rather than replacing them.
As observed by Kluter et al. , this leads to a coherence problem, as the DMA
transfers between main memory and AVS memories bypass the cache hierarchy,
and no mechanism exists to ensure coherence between the AVS memory and the
data cache. They corrected the problem using a hardware coherence protocol;
however, the area overhead of the DMA controller and the coherence protocol
were significant. Additionally, coherence messages transmitted on the bus con-
sume energy, and may increase the latency of accesses to off-chip memory.
Under this scheme, the compiler inserts Speculative DMA transfer instructions
to move data from main memory into the AVS memory before an ISE that
accesses the latter may execute. Each line of the AVS memory is augmented
with valid and dirty bits, similar in principle to a cache, but without the tag
arrays. These bits are required to integrate the AVS memory into the hardware
coherence protocol; additionally, the valid bits facilitate the speculative aspect
of DMA, by suppressing transfers when data in the AVS memory is up-to-date.
The data cache and AVS memory snoop bus transactions. A write to data
in the AVS (cache) invalidates a copy of the data that may exist in the cache
(AVS). If the processor reads invalid data from the cache, the coherence protocol
retrieves the valid copy from the AVS memory. Speculative DMA transfers from
main memory into the AVS memory request the write-back o a dirty copy of
the data that may exist in the cache. When a DMA transfer overwrites valid
and dirty data in the AVS memory, the coherence protocol ensures that the
valid data is written back to main memory. When a DMA transfer overwrites
valid and dirty data in the AVS memory, the coherence protocol ensures that
the valid data is written back to main memory, eliminating the need for explicit
DMA transfer instructions to remove data from the AVS memory.
Virtual Ways, presented here, is a lower-cost solution to the coherence prob-
lem. Unlike Speculative DMA, Virtual Ways uses a relaxed form of inclusion,
in which the AVS memory always contains a subset of the data in the cache;
however, ISE writes to the AVS memory employ a write-back policy that only
updates the copy in the data cache when the processor, later, tries to read the
data from the cache. Speculative DMA, in contrast, does not enforce inclusion.
Under Virtual Ways, data is loaded into the AVS memory using prefetch in-
structions, which eliminates the DMA engine. There is no hardware coherence
protocol, which eliminates both the hardware overhead (i.e., duplicated tags)
and the performance and energy overhead due to snooping and coherence traffic
on the system bus. This improves both system performance and energy con-
Way Stealing is another solution to the coherence problem for AVS-enhanced
ISEs . The data cache is modified so that each way can be accessed as a
compiler controlled memory, and all of the ways can be read or written in parallel.
Each way, however, is a single-ported memory, which can limit the attainable
speedup. For JPEG compression, for example, our ideal AVS memory is a 64-
entry register file with 8 read and 8 write ports; the large number of read and
write ports are feasible for such a small structure, but are not generally scalable
130T. Kluter et al.
ISEProcessor 2 Processor 1
(a) State-of-the-art Automatic In-
struction Set Extension algorithms
provide high bandwidth to the ISE
logic by adding Architecturally Vis-
ible Storage; however, they require
extensive hardware added to a stan-
dard processor pipeline to guarantee
memory coherence [3,9].
ISE Processor 2 Processor 1
(b) Virtual Ways, the contribution
of this work, puts the AVS on top
of the data cache and extends the
cache controller state machine to en-
force coherence. This approach re-
moves the separate bus interface of
the AVS and the need for a coher-
ence protocol in single processor sys-
Fig.1. The difference between providing coherence in Speculative-DMA and Virtual
Ways. In Speculative DMA the AVS is placed at the same level as the L1-caches. In
Virtual Ways, on the other hand, the AVS is placed above the L1-caches and coherence
is provided by inclusion.
for a larger number of entries. Under Way Stealing, the read and write operations
to each stolen way must be serialized due to th small number of ports, which
limits the maximum attainable speedup.
Historically, a single cache based processor system allows for a maximum of two
copies of a given data structure in the system. One copy is always in main mem-
ory and one can be in the cache. In an n-way set associative cache, the location
of a datum within the cache is indicated by the tag arrays and the associated
status bits. The cache state-machine keeps track of the datum by updating the
tag and state arrays accordingly. Any memory element in the system that is not
covered by the tag and state arrays of the cache may exhibit coherence problems.
This is precisely what occurs when AVS is introduced to an extensible processor
without some form of coherence. The most recent copy of a particular datum
may reside in the AVS, rather than the cache. Main memory, therefore, is liable
to load an invalid copy of the same datum into the cache, unless it first updates
the value from the AVS.
This is the classic problem of cache coherence; the fact that the AVS is not
actually a cache does not, in principle, alter the problem; however, it does offer
the possibility of a novel lightweight solution that is considerably less costly than
a full-blown coherence protocol, which in the past has been used for multipro-
cessor systems. Our solution, which we call Virtual Ways, is to treat the AVS as
Virtual Ways: Efficient Coherence for Architecturally Visible Storage131
Fig.2. The AVS is segmented in chunks the size of a cache line and the state is
maintained for each segment separately. The tag consists of the start address and end
address (length) of the AVS. For optimal performance care must be taken to avoid false
sharing between neighboring data structures.
an additional way of the cache with respect to coherence. ISEs still access the
AVS memory like a scratchpad under control of the compiler. The tag associated
with the AVS memory, which is only used to ensure coherence, is implemented
inside the cache. This way, ISE accesses to the AVS memory bypass the tag,
which saves energy on each lookup. The cache controller is aware of the status
of the data residing in the AVS due to its tag, and takes appropriate actions to
ensure coherence. Virtual Ways can ensure coherence between an L1 cache and
an AVS memory in a uniprocessor system.
For easier integration into the cache some adaptations are needed in compari-
son to scratchpad memories. The memory for the data structure held in the AVS
memory is padded to a multiple of the size of a cache line. As the data structure
to be loaded in the AVS is not necessarily aligned on a cache line boundary, the
AVS must hold one additional cache line in order to accommodate all possible
alignments. For optimal performance, an AVS-aware compiler could align data
structures to avoid false sharing. For example, suppose that one data structure
ends near the beginning of a cache line, and another data structure starts some-
where later on the same line. A write to a location in either data structure that
resides on the cache line will invalidate the entire line, including a portion of the
other data structure. This could, in principle, create unnecessary data transfers
between the cache memories and the AVS.
Figure 2 illustrates the memory structure used to implement an AVS as a
Virtual Way. Two bits per segment are required: one bit determines whether the
segment is valid, and the second bit determines whether the copy in the AVS is
exclusive. One set of tags for the AVS indicates the starting and ending addresses
of the data structure stored in the AVS. This set of tags is used to determine
if a CPU access issued to the cache is within the region contained within the
AVS. The set of tags and the state bits permit the cache controller to determine
where the most recent copy of the requested datum resides.
132T. Kluter et al.
Fig.3. Transition digram for each segment of the AVS; each segment can be in one of
three states Invalid, Valid, or Exclusive. Associated ISEs execute while all segments are
either valid or exclusive. ISE writes to an AVS segment cause it to become exclusive.
An exclusive segment must be written back to the cache when the segment transitions
into another state.
An ISE enhanced with an AVS can only execute when all segments are valid, as
all accesses to the AVS must hit. We do not impose any restrictions on the ISE’s
access patterns within the AVS, beyond the requirement that the data reside
in the AVS before the ISE begins to execute. Specialized prefetch instructions
are used to load data into the AVS and update the tag before the ISE can
execute. Similar to caches, data eviction from the AVS is achieved via lazy write
back; however, an AVS-flush instruction is also available. If the data is accessed
through a normal software instruction, the cache controller, which maintains
coherence, will copy the data into the cache, and invalidate the data in the AVS
if it is overwritten. In our experiments, we did not use the AVS-flush operation.
Our expectation is that the AVS flush operation would only be used to facilitate
context switching; our evaluation platform is application-specific, so we do not
employ multiple processes and context switching does not occur.
3.1 AVS Segment States
Each segment of the AVS can be in one of three states. These are:
1. Invalid State: the initial state of the AVS, in which no segment contains
valid data. This occurs when the processor is first powered up, or if the
AVS contains a copy of a data structure that is not the most recent, i.e., a
separate copy, either in the cache or main memory, has been modified, while
the copy residing in the AVS memory has not been updated.
2. Valid State: a segment of the AVS contains the most recent copy of a data
structure. Valid copies of the same line also exist in the cache.
3. Exclusive State: a segment of the AVS contains the most recent copy of a
data structure. The copy in the cache, if any, is dead.
Figure 3 depicts the state machine for one segment of an AVS. Dashed arrows
indicate the transitions where the data must be written back to the cache.
Virtual Ways: Efficient Coherence for Architecturally Visible Storage133
Here, we describe the basic actions of the prefetch instruction, which must com-
plete before an ISE can access the AVS. Here, we define an AVS region to be a
set of m segments, each of which is equal to the size of a cache line. There are
two general cases to consider:
1. AVS Region Match: This occurs if the address of the requested data matches
a segment contained within the AVS. If the state of the segment is valid or
exclusive, then the most recent copy of the data already exists in the AVS; the
data must be loaded into the AVS only if the state is invalid. If a valid copy
of the data exists in another way of the cache, then it can be loaded directly
into the AVS, bypassing the bus; otherwise, the data is loaded from main
memory and is written to the cache and AVS concurrently. See Figure 4 (e)
for a prefetch operation that reloads only one segment.
2. AVS Region Mismatch: This occurs if the address of the requested data
does not match a segment contained within the AVS. If one of the segments
contained within the AVS is currently exclusive, then it must be written
back to the cache/main memory so that the most recent copy of the data
is not lost. Afterwards, all segments are marked invalid and the start and
stop tags are updated for the new data structure. The load operation then
proceeds as described above, with a region match and the AVS segments in
an invalid state. See Figure 4 (f) for the case where the AVS is written back
before it is loaded with a new data structure.
The region matching behavior enforces an inclusive, write-through policy. In-
clusion is maintained, because the lines in the AVS are a subset of the lines in
the cache. This is a relaxed form of inclusion, however, because ISE writes that
modify an AVS segment do not modify the corresponding line in the cache. The
policy is write through, in the sense that prefetch instructions write “through”
the cache directly to the AVS.
3.3 Maintaining Coherence After the ISE Executes
We assume that the data has been prefetched into the AVS, as described in the
preceding section. When an ISE executes, it may modify the data structure in the
AVS. If the data is modified, then at least one line is left in the exclusive state.
After the ISE executes, control returns to the CPU. The data in the AVS will
either be written back upon request, or as dictated by coherence requirements.
The correct action to take by a software load or store instruction depends on
the state of the segment.
1. Invalid State: An invalid segment can be ignored; the data at the requested
address resides in the cache or main memory.
2. Valid State: Here, the AVS contains valid data that was not modified by the
ISE. A valid copy of the data may also exist in the cache. For a read access,
either valid copy of the data can be returned. Writes are somewhat more
134T. Kluter et al.
Fig.4. This figure shows some lines of the cache and the corresponding segments in the
AVS together with associated state bits during a typical AVS scenario. The AVS starts
up in invalid state (a) and is then preloaded with a data structure (b) and transitions
to valid state. Execution of the ISE will modify the data structure turning on (some of)
its exclusive bits (c). On a CPU access the data is copied back to the cache and, on a
write access, invalidated in the AVS (d). A prefetch instruction for the same structure
will restore it to the AVS (e). A prefetch instruction for another structure will write
back all exclusive lines and load the requested structure (f).
complex, as coherence must be maintained between the cache and the AVS.
One possibility is to employ a write-through policy that updates the data in
both the AVS and the cache; a second alternative is to update the data in
the cache and invalidate the data in the AVS. We have opted for the latter
option, because a pipelined write-through could potentially cause a mem-
ory consistence problem between the data cache and the AVS. A memory
consistence problem occurs when a read of data does not return the latest
value written to it. This situation can occur with a pipelined write-through.
Applying a write-through without pipelining would drastically impact the
processor’s critical path.
3. Exclusive State: In this case, only the AVS contains the most recent copy of
the data, and this copy must be written back to the cache before the access
can complete; the corresponding line in the cache is marked as dirty, and
the AVS segment reverts to the valid state, as the data in the AVS is no
longer exclusive. Figure 4 (d) depicts the case of a CPU write access when
the corresponding AVS segment is in exclusive state.
3.4Multiple AVS Memories
The preceding discussion assumes that there is one AVS memory. In principle,
an ISE may access multiple data structures, and writes to both may benefit
from parallel execution. In this case, we would want to instantiate multiple
AVS memories: one per data structure. To facilitate this change, we require an
additional tag and state bits for each AVS that must be checked to maintain
Virtual Ways: Efficient Coherence for Architecturally Visible Storage135
The compiler can avoid inter-AVS transfers by guaranteeing that memory
regions loaded in distinct AVS memories will never overlap. In the most general
case, pointer analysis is undecidable. As described by Biswas et al. , only data
structures that have been disambiguated can be moved into an AVS memory.
Although this approach is conservative, it is necessary to ensure correctness when
compiling languages such as C/C++ that permit arbitrary pointer arithmetic.
Our experimental platform is an internally-developed FPGA-based soft proces-
sor that implements the OpenRISC instruction set. We modified the data cache
implementation to account for Speculative DMA  and Virtual Ways. Our
multi-processor platform allows us to emulate from one to seven OpenRISC pro-
cessors. The platform has software-configurable 16kB instruction caches and
software-configurable 16kB data caches with a choice of MSI-states, MESI-
states, or disabled hardware coherence protocol. Our implementation of Specu-
lative DMA uses the MESI-states protocol in our experiments. Our implementa-
tion of Virtual Ways eliminates the DMA controller, as data is brought into the
AVS memory through the data cache interface. The only other hardware mod-
ification was to augment the cache state machine as described in the preceding
Mimosys Clarity, a compiler that uses the algorithm proposed by Biswas ,
identified the ISEs and generated the VHDL implementations of the ISE logic.
We modified the AVS memory to support Speculative DMA through a DMA
interface and Virtual Ways through the data cache interface; the appropriate
interface is selected via software control. A system deployed in the real world
would support one option or the other, but not both.
Our goal is to demonstrate that Virtual Ways offers a comparable speedup to
Speculative DMA, but at a significantly reduced hardware and energy cost. We
took the EEMBC consumer V2 testbench suite and performed an ISE identifica-
tion on the unmodified source code by taking the first dataset of each algorithm
as test case. It has to be noted at this point that Mimosys Clarity does not
implement the opportunistic Speculative DMA as proposed by Kluter et al. .
All the C-code has been cross-compiled using a gcc 3.4.4 toolchain based on
“newlib” for the OpenRISC.
5 Experimental Results
To perform a comparison between the different methods, we performed a de-
sign space exploration of all algorithms on a non-ISE enhanced processor. We
varied the size and associativity of both the instruction and data caches. The
configuration with the best energy-performance product for a given algorithm
and dataset is chosen as reference for comparison. We performed a similar design
space exploration for the processor augmented with larger AVS-enhanced ISEs,
using both Speculative DMA and Virtual Ways to ensure coherence.
136T. Kluter et al.
Fig.5. Design space exploration of the CJPEGV2 dataset 1 compression algorithm for
the different architectural versions.
The result of the design space exploration for the CJPEGV2 testbench using
the first data set is plotted in Figure 5. Both Speculative DMA and Virtual Ways
achieved greater speedups than the original code across all cache configurations.
Many, but not all, configurations achieved greater reductions in energy when
Speculative DMA or Virtual Ways were used. Except for the reference cache
configuration, the figure does not indicate which Speculative DMA and Virtual
Way data points correspond to the same configuration; the general trend, how-
ever, appears to be that Virtual Ways achieve marginal better performance with
a noticeable reduction in energy compared to Speculative DMA.
Figure 6 shows the energy and performance plots of the EEMBC consumer
version 2 benchmark suite. For all benchmarks Virtual Ways outperforms the
state-of-the-art while consuming significantly less energy. There are two observa-
tions to be made: (1) for the AES algorithm both Speculative DMA and Virtual
Ways perform equally with an identical energy footprint. The reason lies in the
detection of two AVS memories that contain read-only data structures; therefore,
both methods do not have to infer coherence traffic, and (2) for the MPEG2 ENC
both methods provide better performance at a significant energy cost when com-
pared to the baseline. The increase in energy lies in the access pattern of the
detected AVS. In the MPEG2 ENC benchmark a temporary buffer of the size
of 8 × 8 16-bit integers is used to perform a 64-point Discrete Cosine Trans-
form (DCT). The DCT is selected as potential ISE, and the buffer is placed in
an AVS. Due to the algorithm the buffer is moved forth and back between the
AVS and the data cache consuming significant energy. In case of execution on a
non-ISE enhanced processor the buffer is never evicted from the data-cache due
to the Least Recently Used (LRU) replacement policy. To explain why the other
algorithms do not suffer similarly, we compare the data points corresponding to
the reference cache configuration of the CJPEGV2 algorithm using dataset 1 in
Virtual Ways: Efficient Coherence for Architecturally Visible Storage137
AES MPEG2_DECMPEG2_ENC CJPEGV2
DS2DS1DS3DS4DS5 DS6 DS7DS1DS1DS2DS2 DS3DS3DS4 DS4DS5DS5DS1
Fig.6. Performance and energy results for four EEMBC benchmarks. Each of the al-
gorithms, except the AES, contains five to seven different datasets (DSx). The baseline
is the cache configuration that provides the best energy-performance product when
running on a non-ISE enhanced processor. Overall, Virtual Ways provides similar to
more performance with significant reduced energy consumption when compared to
Figure 7 shows the performance and energy breakdown for the four differ-
ent kernels of the CJPEGV2 algorithm for the reference cache configuration.
Similarly to the MPEG2 ENC benchmark the DCT kernel is the only kernel
containing a custom instruction with an AVS. One would expect to observe two
different scenarios: (1) upon entering the DCT kernel, the data has to be copied
to the AVS, before the custom instruction can start processing the data, and (2)
after leaving the DCT kernel the data has to gradually move back to the data
cache for the processor to be able to process it in the quantization kernel.
Looking into the copying of the data structure into the AVS, Figure 7 shows no
distinct differences between Speculative DMA and Virtual Ways in terms of per-
formance or energy consumption, contrary our observation for the MPEG2 ENC
benchmark. The reason for this lies in the calculation pattern of the color space
conversion. The color space conversion processes a “band” of 1024 pixels, 8 rows
at a time. As this “band” corresponds to a memory size of 24kB, it cannot fit
in the data cache entirely, and therefore will evict parts of the processed data.
By the time the DCT kernel starts processing, the data required in the AVS
is no longer present in the data cache; therefore, no coherence problem exists
and both Speculative DMA and Virtual Ways need to prefetch the data from
main memory. As this process affects both methods, both architectures perform
equally and consume about the same amount of energy in this particular case.
138T. Kluter et al.
values (many 0)
Ref S−DMA VW
Ref S−DMA VW
Relative energy consumption
Fig.7. Left: Schematic diagram of the kernels of the CJPEGV2 compression algorithm.
Right: Performance and energy consumption broken down into the different kernels as
shown on the left for the baseline (Ref), Speculative DMA (S-DMA), and Virtual Ways
Figure 7 shows distinct differences for the data eviction process from the AVS.
Where for Speculative DMA the energy consumption in the quantization kernel is
high (4.4× the energy consumed by the non-ISE enhanced architecture), Virtual
Ways expends a comparable amount of energy as the software implementation.
The reason for this is that the data structure in the AVS has been modified by
an ISE in the DCT kernel and is then directly used in the quantization kernel.
In this case, a coherence problem exists between the AVS and the data cache.
In Speculative DMA the coherence protocol will move the data structure back
from the AVS to both the data cache and main memory, which includes expensive
bus transfers; this consumes a significant amount of energy. In contrast, Virtual
Ways simply copies the data directly from the AVS segments to the cache. This
eliminates the need for bus transfers and writes to main memory.
The bus dependency of the Speculative DMA coherence mechanism is an un-
certainty. Due to the well known memory wall problem the processor normally
runs at higher clock frequencies than the external memory. For all of the pre-
ceding experiments, we assumed memory and processor frequencies of 100MHz,
which is a favorable situation for Speculative DMA. Increasing the processor
clock frequency can influence the operation of Speculative DMA in the bench-
mark, as shown in Figure 8(a); Figure 8(a) also shows that the performance of
Virtual Ways is less dependent on the difference between processor and memory
To compare the area of Virtual Ways and Speculative DMA, we implemented
both data caches, including AVS memories, in a 90 nm standard-cell technology,
along with a baseline cache without an AVS; we did not synthesize instruction
caches, the processor, or the ISE computational logic. The results are depicted
in Figure 8(b), which shows that Virtual Ways increases the area of the baseline
cache by 9%, while Speculative DMA increases the area by 29%.
Virtual Ways: Efficient Coherence for Architecturally Visible Storage139
100 200 300 400 500 600
Processor frequency [MHz]
(a) Influence of the processor frequency with
respect to the external memory frequency for
the execution of the CJPEGV2 benchmark.
Ref S−DMA VW
Relative data cache area
(b) Area overhead comparison of a
standard data cache (Ref), a Spec-
ulative DMA enhanced data cache
(S-DMA), and a Virtual Ways en-
hanced data cache (VW).
Fig.8. Frequency robustness and Area of Virtual Ways compared to Speculative DMA
Prior work has established that AVS-enhanced ISEs provide a performance im-
provement over ISEs that do not employ AVS; however, the inclusion of AVS in
a processor with caches creates a memory coherence problem. This paper has
introduced Virtual Ways as a low-cost alternative to using a coherence protocol
to maintain this coherence in a single-processor system. Our results show that
a cache enhanced with Virtual Ways consumes less area and energy than Spec-
ulative DMA; additionally, Virtual Ways was shown to be less sensitive than
Speculative DMA to differences in clock frequencies between the processor and
main memory. For these reasons, we believe that Virtual Ways is a much more
attractive solution than Speculative DMA for customizable processors used in
cost and energy-constrained embedded systems.
1. Atasu, K., Mencer, O., Luk, W.,¨Ozturan, C., D¨ unda, G.: Fast custom instruction
identification by convex subgraph enumeration. In: Proceedings of the 19th Inter-
national Conference on Application-specific Systems, Architectures and Processors,
Leuven, Belgium, July 2008, pp. 1–6 (2008)
2. Biswas, P., Choudhary, V., Atasu, K., Pozzi, L., Ienne, P., Dutt, N.: Introduction
of local memory elements in instruction set extensions. In: Proceedings of the 41st
Design Automation Conference, San Diego, Calif., June 2004, pp. 729–734 (2004)
3. Biswas, P., Dutt, N., Pozzi, L., Ienne, P.: Introduction of architecturally visible
storage in instruction set extensions. IEEE Transactions on Computer-Aided De-
sign of Integrated Circuits and Systems CAD-26(3), 435–446 (March 2007)
140T. Kluter et al.
4. Clark, N., Zhong, H., Mahlke, S.: Processor acceleration through automated in-
struction set customisation. In: Proceedings of the 36th Annual International Sym-
posium on Microarchitecture, San Diego, Calif., December 2003, pp. 129–140 (2003)
5. Cong, J., Han, G., Zhang, Z.: Architecture and compiler optimizations for data
bandwidth improvement in configurable embedded processors. IEEE Transactions
on Very Large Scale Integration (VLSI) Systems 14(9), 986–997 (2006)
6. Jayaseelan, R., Liu, H., Mitra, T.: Exploiting forwarding to improve data band-
width of instruction-set extensions. In: Proceedings of the 43rd Design Automation
Conference, San Francisco, Calif., July 2006, pp. 43–48 (2006)
7. Karuri, K., Chattopadhyay, A., Hohenauer, M., Leupers, R., Ascheid, G., Meyr,
H.: Increasing data-bandwidth to instruction-set extensions through register clus-
tering. In: Proceedings of the International Conference on Computer Aided Design,
San Jose, Calif., November 2007, pp. 166–171 (2007)
8. Kluter, T., Brisk, P., Charbon, E., Ienne, P.: Way stealing: Cache-assisted auto-
matic instruction set extensions. In: Proceedings of the 46th Design Automation
Conference, San Francisco, Calif., July 2009, pp. 31–36 (2009)
9. Kluter, T., Brisk, P., Ienne, P., Charbon, E.: Speculative DMA for Architecturally
Visible Storage in Instruction Set Extensions. In: Proceedings of the International
Conference on Hardware/Software Codesign and System Synthesis, Atlanta, Ga.,
October 2008, pp. 243–248 (2008)
10. Pothineni, N., Kumar, A., Paul, K.: Application specific datapath extension with
distributed I/O functional units. In: Proceedings of the 20th International Confer-
ence on VLSI Design, Bangalore, India (January 2007)
11. Pozzi, L., Atasu, K., Ienne, P.: Exact and approximate algorithms for the extension
of embedded processor instruction sets. IEEE Transactions on Computer-Aided
Design of Integrated Circuits and Systems CAD-25(7), 1209–1229 (2006)
12. Pozzi, L., Ienne, P.: Exploiting pipelining to relax register-file port constraints
of instruction-set extensions. In: Proceedings of the International Conference on
Compilers, Architectures, and Synthesis for Embedded Systems, San Francisco,
Calif., September 2005, pp. 2–10 (2005)
13. Steinke, S., Wehmeyer, L., Lee, B.-S., Marwedel, P.: Assigning program and data
objects to scratchpad for energy reduction. In: Proceedings of the Design, Automa-
tion and Test in Europe Conference and Exhibition, Paris (March 2002)
14. Verma, A.K., Brisk, P., Ienne, P.: Rethinking custom ISE identification: A new
processor-agnostic method. In: Proceedings of the International Conference on
Compilers, Architectures, and Synthesis for Embedded Systems, Salzburg, Septem-
ber 2007, pp. 125–134 (2007)
15. Verma, A.K., Brisk, P., Ienne, P.: Fast, quasi-optimal, and pipelined instruction-
set extensions. In: Proceedings of the Asia and South Pacific Design Automation
Conference, Seoul, Korea, January 2008, pp. 334–339 (2008)