Reiner W. Hartenstein
`Herbert Griinbacher (Eds.)
`
`Field-Program.m.able
`Logic and Applications
`
`The Roadmap to Reconfigurable Computing
`
`10th International Conference, FPL 2000
`Villach, Austria, August 27-30, 2000
`· Proceedings
`
`Springer
`
`Petitioners Amazon
`Ex. 1003, p. 1
`
`

`

`Series Editors
`
`Gerhard Goos, Karlsruhe University, Germany
`Juris Hartmanis, Cornell University, NY, USA
`Jan van Leeuwen, Utrecht University, The Netherlands
`
`Volume Editors
`
`Reiner W. Hartenstein
`University of Kaiserslautem, Computer Science Department
`P. 0. Box. 30 49, 67653 Kaiserslautern, Germany
`E-mail: hartenst@rhrk.uni-kl.de
`Herbert Griinbacher
`Carinthia Tech Institute
`Richard-Wagner-Str. 19, 9500 Villach, Austria
`E-mail: hg@cti.ac.at
`
`Cataloging-in-Publication Data applied for
`
`Die Deutsche Bibliothek - CIP-Einheitsaufnahrne
`
`Field programmable logic and applications : the roadmap to
`reconfigurable computing ; 10th international conference ; proceedings
`/ FPL 2000, Villach, Austria, August 27 - 30, 2000. Reiner W.
`Hartenstein; Herbert Griinbacher (ed.). - Berlin ; Heidelberg; New
`York ; Barcelona ; Hong Kong ; London ; Milan ; Paris ; Singapore ;
`·
`Tokyo : Springer, 2000
`(Lecture notes in computer science ; Vol. l 896)
`
`-lSBN ,~1ifrfi~9We'ndt Library
`
`University of Wisconsin-Madison
`215 N. Rand al l Avenue
`Madison, WI 53706-1688
`
`CR Subject Classification (1998): B.6-7, J.6
`
`ISSN 0302-9743
`ISBN 3-540-67899-9 Springer-Verlag Berlin Heidelberg New York
`
`This work is subject to copyright. All rights are reserved, whether the whole or part of the material is
`concerned, specifically the rights of translation, reprinting, re-use of illustrations, recitation, broadcasting,
`reproduction on microfilms or in any other way, and storage in data banks. Duplication of this publication
`or parts thereof is permitted only under the provisions of the German Copyright Law of September 9, 1965,
`in its current version, and permission for use must always be obtained from Springer-Verlag. Violations are
`liable for prosecution under the German Copyright Law.
`
`Springer-Verlag Berlin Heidelberg New York
`a member of BertelsmannSpringer Science+Business Media GmbH
`© Springer-Verlag Berlin Heidelberg 2000
`Printed in Germany
`
`Typesetting: Camera-ready by author, data conversion by Steingraber Satztechnik GmbH, Heidelberg
`Printed on acid-free paper
`SPIN 10722573
`06/3142
`5 4 3 2 I 0
`
`Petitioners Amazon
`Ex. 1003, p. 2
`
`

`

`Memory Access Schemes for Configurable Processors
`
`Bolger Lange and Andreas Koch
`
`Tech. Univ. Braunschweig (E.I.S.), GauBstr. 11, D-38106 Braunschweig, Germany
`lange,koch@eis.cs.tu-bs.de
`
`Abstract. This work discusses the Memory Architecture for Reconfigurable Com(cid:173)
`puters (MARC), a scalable, device-independent memory interface that supports
`both irregular (via configurable caches) and regular accesses (via pre-fetching
`stream buffers). By hiding specifics behind a consistent abstract interface, it is
`suitable as a target environment for automatic hardware compilation.
`
`1 Introduction
`
`Reconfigurable compute elements can achieve considerable performance gains over
`standard CPUs [1] [2] [3] [4]. In practice, these configurable elements are often combined
`with a conventional processor, which provides the control and 1/0 services that are
`implemented more efficiently in fixed logic. Recent single-chip architectures following
`this approach include NAPA [5], GARP [6], OneChip [7], OneChip98 [8], Triscend
`ES [9], and Altera Excalibur [10]. Board-level configurable processors either include a
`dedicated CPU [11] [12] or rely on the host CPU for support [13] [14] .
`Design tools targeting one of these hybrid systems such as GarpCC [15], Nimble
`[16] or Napa-C [17] have to deal with software and hardware issues separately as well as
`with the creation of interfaces between these parts. On the software side, basic services
`such as 1/0 and memory management are often provided by an operating system of
`some kind. This can range from a full-scale general-purpose OS over more specialized
`real-time embedded OSes down to tiny kernels offering only a limited set of functions
`tailored to a very specific class of applications. Usually, a suitable OS is either readily
`available on the target platform, or can be ported to it with relative ease.
`This level of support is unfortunately not present on the hardware side of the hybrid
`computer. Since no standard environment is available for even the most primitive tasks
`such as efficient memory access or communication with the host, the research and de(cid:173)
`velopment of new design tools often requires considerable effort to provide a reliable
`environment into which the newly-created hardware can be embedded. This environment
`is sometimes called a wrapper around the custom datapath. It goes beyond a simple as(cid:173)
`signment of chip pads to memory pins. Instead, a structure of on-chip busses and access
`protocols to various resources (e.g., memory, the conventional processor, etc) must be
`defined and implemented.
`In this paper, we present our work on the Memory Architecture for Reconfigurable
`Computers (MARC). It can act as a "hardware target" for a variety of hybrid compil(cid:173)
`ers, analogously to a software target for conventional compilers. Before describing its
`specifics, we will justify our design decisions by giving a brief overview of current con(cid:173)
`figurable architectures and showing the custom hardware architectures created by some
`hybrid compilers.
`
`R.W. Hartenstein and H. Griinbacher (Eds.) FPL 2000, LNCS 1896, pp. 615-625, 2000.
`© Springer-Verlag Berlin Heidelberg 2000
`
`Petitioners Amazon
`Ex. 1003, p. 3
`
`

`

`616
`
`H. Lange and A. Koch
`
`2 Hybrid Processors
`
`Static and reconfigurable compute elements may be combined in many ways. The degree
`of integration can range from individual reconfigurable function units (e.g., 0neChip
`[7]) to an entirely separate coprocessor attached to a peripheral bus (e.g., SPLASH [4],
`SPARXIL [18)).
`
`::E
`<
`a:
`Q
`
`a. :c
`0
`0 en
`en
`e
`GI
`:2 .s >,
`
`D.
`
`::c
`
`1/0 Bus
`
`Figure 1. Single-chip hybrid processor
`
`Figure 1 sketches the architecture of a single-chip hybrid processor that combines
`fixed (CPU) and reconfigurable (RC) compute units behind a common cache (D$).
`Such an architecture was proposed, e.g., for GARP [6) and NAPA [5]. It offers very
`high bandwidth, low latency, and cache coherency between the CPU and the RC when
`accessing the shared DRAM.
`
`Fixed Processor Chip
`
`Reconfigurable Array
`
`DRAM
`
`SRAM
`
`Figure 2. Hybrid processor emulated by multi-chip system
`
`The board-level systems more common today use an architecture similar to Figure
`2 . Here, a conventional CPU is attached by a bus interface unit (BIU) to a system-wide
`1/0 bus (e.g., SBus [18) or PCI [11) [12)). Another BIU connects the RC to the 1/0
`bus. Due to the high communication latencies over the 1/0 bus, the RC is often attached
`directly to a limited amount of dedicated memory (commonly a few KB to a few MB of
`
`Petitioners Amazon
`Ex. 1003, p. 4
`
`

`

`Memory Access Schemes for Configurable Processors
`
`617
`
`SRAM). In some systems, the RC has access to the main DRAM by using the 1/0 bus as
`a master to contact the CPU memory controller (MEMC). With this capability, the CPU
`and the RC are sharing a logically homogeneous address space: Pointers in the CPU
`main memory can be freely exchanged between software on the CPU and hardware in
`the RC.
`Operation
`ZBT SRAM read
`ZBT SRAM write
`PCI read
`PCI write
`
`Cycles
`4
`4
`46-47
`10
`
`Table 1 shows the latencies measured on [12) for the
`RC accessing data residing in local Zero-Bus Turnaround
`(ZBT) SRAM (latched in the FPGA 1/0 blocks) and in
`main DRAM (via the PCI bus). In both cases, one word
`per cycle is transferred after the initial latency.
`It is obvious from these numbers that any useful wrap(cid:173)
`per must be able to deal efficiently with access to high la(cid:173)
`tency memories. This problem, colloquially known as the
`"memory bottleneck", has already been tackled for con(cid:173)
`ventional processors using memory hierarchies (multiple cache levels) combined with
`techniques such as pre-fetching and streaming to improve their performance. As we will
`see later, these approaches are also applicable to reconfigurable systems.
`
`Table 1. Data access laten-
`cies (single word transfers)
`
`3 Reconfigurable Datapaths
`
`The structure of the compute elements implemented on the RC is defined either manually
`or by automatic tools. A common architecture [6] [16] [18] is shown in Figure 3.
`
`Hardware Operators
`
`~
`0
`E
`CII
`E
`
`1 (.)
`
`a:
`
`I •
`
`.. • ..
`
`Figure 3. Common RC datapath architecture
`
`The datapath is formed by a number of hardware operators, often created using
`module generators, which are placed in a regular fashion. While the linear placement
`shown in the figure is often used in practice, more complicated layouts are of course
`possible. All hardware operators are connected to a central datapath controller that
`orchestrates their execution.
`In this paper, we focus on the interface blocks attaching the datapath to the rest of the
`system. They allow communication with the CPU and main memory using the system
`bus or access to the local RC RAM. The interface blocks themselves are accessed by the
`datapath using a structure of uni- and bidirectional busses that transfer data, addresses,
`and control information.
`
`Petitioners Amazon
`Ex. 1003, p. 5
`
`

`

`618
`
`H. Lange and A. Koch
`
`For manually implemented RC applications, the protocols used here are generally
`developed ad-hoc and heavily influenced by the specific hardware environment targeted.
`(e.g., the data sheets of the actual SRAM chips on a PCB). In practice, they may even
`vary between different applications running on the same hardware (e.g., usage of burst(cid:173)
`modes, fixed access sizes etc.).
`This approach is not applicable for automatic design flows: These tools require pre(cid:173)
`defined access mechanisms to which they strictly adhere for all designs. An example for
`such a well-defined protocol suitable as a target for automatic compilation is employed on
`GARP [6], a single-chip hybrid processor architecture. It includes standardized protocols
`for sending and retrieving data to/from the RC using specialized CPU instructions and
`supported by dedicated decoding logic in silicon. Memory requests are routed over a
`single address and four data busses that can supply up to four words per cycle for regular
`(streaming) accesses.
`None of these capabilities is available when using off-the-shelf silicon to implement
`the RC. Instead, each user is faced with implementing the required access infrastructure
`anew.
`
`4 MARC
`
`Our goal was to learn from these past experiences and develop a single, scalable, and
`portable memory interface scheme fpr reconfigurable datapaths. MARC strives to be
`applicable for both single-chip and board-level systems, and to hide the intricacies of
`different memory systems from the datapath. Figure 4 shows an overview of this
`architecture.
`
`Back-Ends
`
`Front-Ends
`
`RC Array
`
`BIU
`
`1/0 Bus
`
`Figure 4. MARC architecture
`
`Petitioners Amazon
`Ex. 1003, p. 6
`
`

`

`Memory Access Schemes for Configurable Processors
`
`619
`
`Using MARC, the datapath accesses memory through abstract front-end interfaces.
`Currently, we support two front-ends specialized for different access patterns: Caching
`ports provide for efficient handling of irregular accesses. Streaming ports offer a non(cid:173)
`unit stride access to regular data structures (such as matrices or images) and perform
`address generation automatically. In both cases, data is pre-fetched/cached to reduce the
`impact of high latencies (especially for transfers using the I/0 bus). Both ports use stall
`signals to indicate delays in the data transfer (e.g., due to cache miss or stream queue
`refill). A byte-steering logic aligns 8- and 16-bit data on bits 7 :0 and 15:0 of the data
`bus regardless of where the datum occurred in the 32-bit memory or bus words.
`The specifics of hardware memory chips or system bus protocols are implemented
`in various back-end interfaces. E.g., dedicated back-ends encapsulate the mechanisms
`for accessing SRAM or communicating over the PCI bus using the BIU.
`The MARC core is located between front- and back-ends, where it acts as the main
`controller and data switchboard. It performs address decoding and arbitration between
`transfer initiators in the datapath and transfer receivers in the individual memories and
`busses. Logically, it can map an arbitrary number of front-ends to an arbitrary number
`of back-ends. In practice, though, the number of resources managed is of course limited
`by the finite FPGA capacity. Furthermore, the probability of conflicts between initiators
`increases when they share a smaller number of back-ends. However, the behavior visible
`to the datapath remains identical: The heterogeneous hardware resources handled by the
`back-ends are mapped into a homogeneous address space and accessed by a common
`protocol.
`
`4.1
`
`Irregular Cached Access
`
`Caching ports are set up to provide read data one cycl~ after an address has been applied,
`and accept one write datum/address per clock cycle. If this is not possible (e.g., a cache
`miss occurs), the stall signal is asserted for the affected port, stopping the initiator.
`When the stall signal is de-asserted, data that was "in-flight" due to a previous request
`will remain valid to allow the initiator to restart cleanly.
`Table 3( a) describes the interface to a caching port. The architecture currently allows
`for 32-bit data ports, which is the size most relevant when compiling software into hybrid
`solutions. Should the need arise for wider words, the architecture can easily be extended.
`Arbitrary memory ranges ( e.g., memory-mapped I/0 registers) can be marked as non(cid:173)
`cacheable. Accesses to these regions will then bypass the cache. Furthermore, since all
`of the cache machinery is implemented in configurable logic, cache port characteristics
`such as number of cache lines and cache line length can be adapted to the needs of the
`application. As discussed in [19], this can result in a 3% to 10% speed-up over using a
`single cache configuration for all applications.
`
`4.2 Regular Streamed Access
`
`Streaming ports transfer a number of data words from or to a memory area without the
`need for the datapath to generate addresses. After setting the parameters of the transfer
`(by switching the port into a "load parameter" mode), the port presents/accepts one data
`item per clock cycle until it has to refill or flush its internal FIFOs. In that case, the stall
`
`Petitioners Amazon
`Ex. 1003, p. 7
`
`

`

`620
`
`H. Lange and A. Koch
`
`Table 2. Port interfaces
`
`Signal Kind Function
`Addr
`in Address.
`Data in/out Data item.
`Width
`in 8, 16, 32-bit access.
`Stall out Asserted on cache miss.
`OE
`in Output enable
`in Write enable
`WE
`Flush
`in Flush cache.
`
`(a) Caching port interface
`
`Sig/Reg Kind Function
`reg Start address.
`Addr
`reg Stride (increment).
`Stride
`Width
`reg 8, 16,32-bit access.
`reg FIFO size.
`Block
`reg Length, transfer.
`Count
`reg Read or write.
`R/W
`Data
`i/o Data item.
`out Wait, FIFO flush/refill .
`Stall
`Hold
`in Pause data flow.
`out End of stream reached.
`EOS
`in Accept new parameters.
`Load
`
`(b) Streaming port interface
`
`signal stops the initiator using the port. When the FIFO becomes ready again, the stall
`signal is de-asserted and the transfer continues. The datapath can pause the transfer by
`asserting the hold signal. As before, our current implementation calls for a 32-bit wide
`data bus. Table 3(b) lists the parameter registers and the port interface.
`The 'Block' register plays a crucial role in matching the stream characteristics to
`the specific application requirements. E.g., if the application has to process a very large
`string (such as in DNA matching), it makes sense for the datapath to request a large
`block size. The longer start-up delay (for the buffer to be filled) is amortized over the
`long run-time of the algorithm. For smaller amounts of data (e.g., part of a matrix row for
`blocking matrix multiplication), it makes much more sense to pre-fetch only the precise
`amount of data required. [20) suggests compile-time algorithms to estimate the FIFO
`depth to use.
`The cache is bypassed by the streaming ports in order to avoid cache pollution.
`However, since logic guaranteeing the consistency between caches and streams for arbi(cid:173)
`trary accesses would be very expensive to implement (especially when non-unit strided
`streams are used), our current design requires that accesses through the caching ports do
`not overlap streamed memory ranges. This restriction must be enforced by the compiler.
`If that is not possible, streaming ports cannot be used. As an alternative, a cache with
`longer cache lines (e.g., 128 bytes), might be used to limit the performance loss due to
`memory latency.
`
`4.3 Multi-threading
`
`Note that all stall or flow-control signals are generated/accepted on a per-port basis.
`This allows true multi-threaded hardware execution where different threads of control
`are assigned to different ports. MARC can accommodate more logical ports on the front(cid:173)
`ends than actually exist physically on the back-ends. For certain applications, this can
`be exploited to allow the compiler to schedule a larger number of memory accesses
`in parallel. The MARC core will resolve any inter-port conflicts (if they occur at all,
`
`Petitioners Amazon
`Ex. 1003, p. 8
`
`

`

`Memory Access Schemes for Configurable Processors
`
`621
`
`see Section 5 ) at run-time. The current implementation uses a round-robin policy, later
`versions might extend this to a priority-based scheme.
`
`4.4 Flexibility
`
`A separate back-end is used for each memory or bus resource. For example, in a system
`with four ZBT SRAM memories, four instances of the ZBT SRAM back-end would be
`instantiated. The back-ends present the same interface as a caching port ( Table 3(a) ) to
`the MARC core. They encapsulate the state and access mechanisms to manage each of
`the physical resources. E.g., a PCI backend might know how to access a PCI BIU and
`initiate a data transfer.
`In this manner, additional back-ends handling more memory banks can be attached
`easily. Analogously, MARC can be adapted to different FPGA technologies. For exam(cid:173)
`ple, on the Xilinx Virtex [24) series of FPGA, the L1 cache of a caching port might
`be implemented using the on-chip memories. On the older XC4000XL series, which
`has only a limited amount of on-chip storage, the cache could be implemented in a
`direct-mapped fashion that has the cache lines placed in external memory.
`
`5 Implementation Issues
`
`Our first MARC implementation is targeting the prototyping environment described in
`[ 12). The architecture details relevant for this paper are show in Figure 5 .
`
`0
`a)
`0 a,
`X
`...J
`Q.
`
`128Kx36b
`
`128Kx36b
`
`128Kx36b
`
`128Kx36b
`
`Figure 5. Architecture of prototype hardware
`
`A SUN microSPARC-Ilep RISC [21) [22) is employed as conventional CPU. The
`RC is composed of a Xilinx Virtex XCVlOOO FPGA [24) .
`
`5.1 Status
`
`At this point, we have implemented and intensively simulated a parameterized Verilog
`model of the MARC Core and back-ends. On the front-end side, caching ports are already
`operational, while streaming ports are still under development. The design is currently
`only partially floorplanned, thus the given performance numbers are preliminary.
`
`Petitioners Amazon
`Ex. 1003, p. 9
`
`

`

`622
`
`H. Lange and A. Koch
`
`5.2 Physical Resources
`
`The RC has four 128Kx36b banks of ZBT SRAM as dedicated memory and can access
`the main memory (64MB DRAM managed by the CPU) over the PCI bus. To this end,
`a PLX 9080 PCI Accelerator [23] is used as BIU that translates the RC bus (i960-like)
`into PCI and back. The MARC core will thus need PLX and ZBT SRAM back-ends.
`All of their instances can operate in parallel.
`
`5.3 MARC Core
`
`The implementation follows the architecture described in Section 4 : An arbitrary number
`of caching and streaming ports can be managed. In this implementation (internally
`relying on the Virtex memories in dual-ported mode), two cache ports are guaranteed to
`operate without conflicts, and three to four cache ports may operate without conflicts. If
`five or more cache ports are in use, a conflict will occur and be resolved by the Arbitration
`unit ( Section 4.3 ). This version of the core currently supports a 24-bit address space
`into which the physical resources are mapped.
`
`5.4 Configurable Cache
`
`We currently provide three cache configurations: 128 lines of 8 words, 64 lines of 16
`words, or 32 lines of 32 words. Non-cacheable areas may be configured at compile
`time (the required comparators are then synthesized directly into specialized logic).
`The datapath can explicitly request a cache flush at any time (e.g., after the end of a
`computation).
`The cache is implemented as a fully associative LI cache. It uses 4KB ofVirtex Block(cid:173)
`SelectRAM to hold the cache lines on-chip and implements write-back and random line
`replacement. The BlockSelectRAM is used in dual-port mode to allow up to two accesses
`to occur in parallel. Conflicts are handled by the MARC Core Arbitration logic. The
`CAMs needed for the associative lookup are composed from SRL16E shift registers as
`suggested in [25] . This allows a single-cycle read (compare and match detection) and 16
`clock cycles to write a new tag into the CAM. Since this operation occurs simultaneously
`with the loading of the cache lines from memory, the CAM latency is completely hidden
`in the longer memory latencies. As this 16-cycle delay would also occur (and could not
`be hidden) when reading the tag, e.g., when writing back a dirty cache line, the tags are
`additionally stored in a conventional memory composed from RAM32x 1 S elements that
`allow single-cycle reading.
`For each caching port, a dedicated CAM bank is used to allow lookups to occur in
`parallel. Each cache line requires 5 4-bit CAMs, thus the per-port CAM area requirements
`range from 160 to 640 4-LUTs. In addition to the CAMs, each cache port includes a
`small state-machine controlling the cache operation for different scenarios (e.g., read
`hit, read miss, write hit, write miss, flush , etc.). The miss penalty Cm in cycles for a clean
`cache line is given by
`
`Cm = 7 + ~e + w + 4,
`where ~e is the data transfer latency for the back-end used ( Table I ) and w is the
`number of 32-bit words per cache line. 7 and 4 are the MARC Core operation startup
`
`Petitioners Amazon
`Ex. 1003, p. 10
`
`

`

`Memory Access Schemes for Configurable Processors
`
`623
`
`and shutdown times in cycles, respectively. For a dirty cache line, an additional w cycles
`are required to write the modified data back.
`For comparison with [19], note that according to [26], the performance of 4KB of
`fully-associative cache is equivalent to that of 8KB of direct-mapped cache.
`
`5.5 Performance and Area
`
`The performance and area requirements of MARC Core, the technology modules and
`two cache ports are shown in Table 3 .
`
`Table 3. Performance and area requirements
`
`RAM32X1Ss
`4LUTs FFs
`Configuration
`12
`2844
`976
`32x32
`26
`4182 1038
`64x16
`56
`7132 1531
`128x8
`XCVlOO0 avail. 24576 24576 (each uses 2 4LUTs)
`
`BlockRAMs Clock
`31 MHz
`8
`30MHz
`8
`29MHz
`8
`-
`32
`
`For the three configuration choices, the area requirements vary between 10%-30%
`of the chip logic capacity. Since all configurations use 4KB of on-chip memory for cache
`line storage, 8 of the 32 512x8b BlockSelectRAMs are required.
`
`5.6 Scalability and Extensibility
`
`As shown in Table 3 , scaling an L1 on-chip above l 28x8 is probably not a wise choice
`given the growing area requirements. However, as already mentioned in Section 4.4 ,
`part of the ZBT SRAM could be used to hold the cache lines of a direct mapped L2
`cache. In this scenario, only the tags would be held inside of the FPGA.
`A sample cache organization for this approach could partition the 24-bit address
`into a 8-bit tag, 12-bit index and 4-bit block offset. The required tag RAM would be
`organized as 4096x8b, and would thus require 8 BlockSelectRAMs (in addition to those
`used for the on-chip L1 cache). The cache would have 4096 lines of 16 words each, and
`would thus require 64K words of the 128K words available in one of the ZBT SRAM
`chips. The cache hit/miss determination could occur in two cycles, with the data arriving
`after another two cycles. A miss going to PCI memory would take 66 cycles to refill a
`clean cache line and deliver the data to the initiator.
`
`6 Related Work
`
`[ 19] gives experimental result for the cache-dependent performance behavior of 6 of the
`8 benchmarks in the SPECint95 suite. Due to the temporal configurability we suggest
`for the MARC caches (adapting cache parameters to applications), they expect a perfor(cid:173)
`mance improvement between 3% to 10% over static caches. [27] describes the use of the
`
`Petitioners Amazon
`Ex. 1003, p. 11
`
`

`

`624
`
`H. Lange and A. Koch
`
`configurable logic in a hybrid processor to either add ~ victim cache or pre-fetch buffers
`to an existing dedicated direct-mapped L1 cache on an per-application basis. They quote
`improvements in L1 miss rate of up to 19%. [28) discusses the addition of 1MB of LI
`cache memory managed by a dedicated cache controller to a configurable processor.
`Another approach proposed in [29) re-maps non-contiguous strided physical addresses
`into contiguous cache line entries. A similar functionality is provided in MARC by the
`pre-fetching of data into the FIFOs of streaming ports. [30) suggests a scheme which
`adds configurable logic to a cache instead of a cache to configurable logic. They hope to
`avoid the memory bottleneck by putting processing (the configurable logic) very close
`to the data. The farthest step with regard to data pre-fetching is suggested in [31), which
`describes a memory system that is cognizant of high-level memory access patterns.
`E.g., once a certain member in a structure is accessed, a set of associated members is
`fetched automatically. However, the automatic generation of the required logic from
`conventional software is not discussed. On the subject of streaming accesses, [20] is an
`exhaustive source. The 'Block' register of our streaming ports was motivated by their
`discussion of overly long startup times for large amounts of pre-fetched data.
`
`7 Summary
`
`We presented an overview of hybrid processor architectures and some memory access
`needs often occurring in applications. For the most commonly used RC components
`( off-the-shelf FPGAs ), we identified ·a lack of support for even the most basic of these
`requirements.
`As a solution, we propose a general-purpose Memory Architecture for Reconfig(cid:173)
`urable Computers that allows device-independent access both for regular (streamed)
`and irregular (cached) patterns. We discussed one real-world implementation of MARC
`on an emulated hybrid processor combining a SPARC CPU with a Virtex FPGA. The
`sample implementation fully supports multi-threaded access to multiple memory banks
`as well as the creation of "virtual" memory ports attached to on-chip cache memory.
`The configurable caches in the current version can reduce the latency from 46 cycles
`(for access to DRAM via PCI) down to a single cycle on a cache hit.
`
`References
`
`1. Amerson, R., "Teramac - Configurable Custom Computing", Proc. IEEE Symp. on FCCMs,
`Napa 1995
`2. Bertin, P., Roncin, D., Vuillemin, J., "Programmable Active Memories: A Performance As(cid:173)
`sessment", Proc. Symp. Research on Integrated Systems, Cambridge (Mass.) 1993
`3. Box, B., "Field-Programmable Gate Array-based Reconfigurable Preprocessor", Proc. IEEE
`Symp. on FCCMs, Napa 1994
`4. Buell, D., Arnold, J., Kleinfelder, W. , "Splash 2 - FPGAs in Custom Computing Machines",
`IEEE Press, 1996
`5. Rupp, C., Landguth, M ., Garverick, et al., "The NAPA Adaptive Processing Architecture",
`Proc. IEEE Symp. on FCCMs, Napa 1998
`6. Hauser, J., Wawrzynek, J., "Garp: A MIPS Processor with a Reconfigurable Coprocessor",
`Proc. IEEE Symp. on FCCMs, Napa 1997
`
`Petitioners Amazon
`Ex. 1003, p. 12
`
`

`

`Memory Access Schemes for Configurable Processors
`
`625
`
`7. Wittig, R., Chow, P., "OneChip: An FPGA Processor with Reconfigurable Logic", Proc. IEEE
`Symp. on FCCMs, Napa 1996
`8. Jacob, J., Chow, P., "Memory Interfacing and Instruction Specification for Reconfigurable
`Processors", Proc. ACM Intl. Symp. on FPGAs, Monterey 1999
`9. Triscend, "Triscend ES CSoC Family", http://www.triscend.com/products/lndexE5.html,
`2000
`10. Altera, "Excalibur Embedded Processor Solutions",
`http://www.altera.com/html/products/excalibur.html, 2000
`11. TSI-Telsys, "ACE2card User's Manual", hardware documentation, 1998
`12. Koch, A., "A Comprehensive Platform for Hardware-Software Co-Design", Proc. Intl. Work-
`shop on Rapid-Systems Prototyping, Paris 2000
`13. Annapolis Microsystems, http://www.annapmicro.com, 2000
`14. Virtual Computer Corp. , http://www.vcc.com, 2000
`15. Callahan, T. , Hauser, J.R., Wawrzynek, J. , "The Garp Architecture and C Compiler", IEEE
`Computer, April 2000
`16. Li, Y., Callahan, T., Darnell, E ., Harr, R., et al., "Hardware-Software Co-Design of Embedded
`Reconfigurable Architectures", Proc. 37th Design Automation Conference, 2000
`17. Gokhale, M.B., Stone, J.M., "NAPA C: Compiling for a Hybrid RISC/FPGA Machine", Proc.
`IEEE Symp. on FCCMs, 1998
`18. Koch, A., Golze, U., "Practical Experiences with the SPARXIL Co-Processor", Proc. Asilo(cid:173)
`mar Conference on Signals, Systems, and Computers, 11/1997
`"Configurable Cache", CMU EE742 course project,
`19. Fung,
`J.M .L.F., Pan,
`J.,
`http://www.ece.cmu.edu/ ee742/proj-s98/fung, 1998
`20. McKee, S.A., "Maximizing Bandwidth for Streamed Computations" , dissertation, U. of Vir(cid:173)
`ginia, School of Engineering and Applied Science, 1995
`21. Sun Microelectronics, "microSPARC-Ilep User's Manual", http://www.sun.com/sparc, 1997
`22. Weaver, D.L., Germond, T., "The SPARC Architecture Manual, Version 8", Prentice-Hall,
`1992
`23 . PLX Technology, "PCI 9080 Data Book", http://www.plxtech.com, 1998
`24. Xilinx, Inc., "Virtex 2.5V Field-Programmable Gate Arrays", http://www.xilinx.com, 1999
`25 . Xilinx, Inc. "Designing Flexible, Fast CAMs with Virtex Family FPGAs", Xilinx Application
`Note 203, 1999
`26. Hennessy, J., Patterson, D., "Computer Architecture: A Quantitative Approach", Morgan(cid:173)
`Kaumann, 1990
`27. Zhong, P., Martonosi, M ., "Using Reconfigurable Hardware to Customize Memory Hierar(cid:173)
`chies'', Proc. SPIE, vol. 2914, 1996
`28 . Kimura, S., Yukishita, M., ltou, Y., et al., "A Hardware/Software Codesign Method for a
`General-Purpose Reconfigurable Co-Processor" , Proc. 5th CODES/CASHE, 1997
`29. Carter, J., Hsieh, W., Stoller, L., et al., "Impulse: Building a Smarter Memory Controller",
`Proc. 5th Intl. Symp. on High. Perf. Comp. Arch. (HPCA) , 1999
`30. Nakkar. M., Harding, J., Schwartz, D. , et al., "Dynamically programmable cache", Proc.
`SPIE, vol. 3526, 1998
`31. Zhang, X. , Dasdan, A., Schulz, M ., et al., "Architectural Adaptation for Application-Specific
`Locality Optimizations", Proc. Intl. Conj on Comp. Design (ICCD), 1997
`
`Petitioners Amazon
`Ex. 1003, p. 13
`
`