H E W L E
`
`' - P A C K A R D
`
`JOURNAL
`
`April 1995
`
`© Copr. 1949-1998 Hewlett-Packard Co.
`
`H E W L E T T
`PACKARD
`
`AMAZON-1021
`8,204,134
`
`

`

`April 1995 Volume 46 • Number 2
`
`H E W L E T T - P A C K A R D
`
`JOURNAL
`
`Articles
`
`6
`12
`23
`
`A Low-Cost, High-Performance PA-RISC Workstation with Built-in Graphics, Multimedia, and
`Networking Capabilities, by Roger A. Pearson
`
`The PA 7100LC Microprocessor: A Case Study of 1C Design Decisions in a Competitive
`Environment, by Mick Bass, Patrick Knebel, David W. Quint, and William L Walker
`
`Design Blanchard, for the PA 7100LC Microprocessor, by Mick Bass, Terry W. Blanchard,
`D. Douglas Josephson, Duncan Weir, and Daniel L Halperin
`
` An I/O McAllister, on a Chip, by Thomas V. Spencer, Frank J. Lettang, Curtis R. McAllister,
`Anthony L Riccio, Joseph F. Orth, and Brian K. Arnold
`
`A X An Workstation, Graphics Accelerator for a Low-Cost Multimedia Workstation, by Paul Martin
`
`HP Color Recovery Technology, by Anthony C. Barkans
`
`R Y T r u e C o l o r
`
`I Real-Time Software MPEG Video Decoder on Multimedia-Enhanced PA 7100LC Processors,
`by Ruby B. Lee, John P. Beck, Joel Lamb, and Kenneth E. Severson
`
`j Overview of the Implementation of the PA 7100LC Multimedia Enhancements
`
`HPTeleShare: Integrating Telephone Capabilities on a Computer Workstation, by S. Paul
`Tucker
`
`72
`
`Caller-ID
`
`/ -4 Call Progress, DTMF Tones, and Tone Detection
`
`Product Design of the Model 712 Workstation and External Peripherals, by Arlen L Hoesner
`
`Editor. Richard P. Dolan • Associate Editor. Charles I. Leath • Publication Production Manager, Susan E. Wright • illustration, Renee D. Pighini
`
`Advisory Thomas Rajeev Badval. Integrated Circuit Business Division. Fort Collins. Colorado » Thomas Beecher. Open Systems Software Division, Cheimsford,
`Massachusettes Chou, Steven Brittenham, Disk Memory Division, Boise, Idaho » Wliliam W. Brown, Integrated Circuit Business Division. Santa Clara, California » Harry Chou,
`Microwave Technology Division, Santa Rosa. California» Rajesh Desai, Commercial Systems Division. Cupertino, California» Kevin G. Ewert, Integrated Systems Division,
`Sunnyvale, Hardcopy • Bernhard Fischer, Boblingen Medical Division, Boblingen, Germany • Douglas Gennetten, Greeley Hardcopy Division, Greeley, Colorado
`» Gary Gordon, Semiconductor Laboratories. Palo Alto, California » Matt J. Marline, Systems Technology Division. Hoseville, California • Kiyoyasu Hiwada, Hachioji Semiconductor
`Test Division. Tokyo, Japan • Bryan Hoog, Lake Stevens Instrument Division, Everett, Washington » Roger L Jungerman, Microwave Technology Division, Santa Rosa.
`California» Ruby B. Lee, Networked Systems Group, Cupertino. California* Alfred Maute. Waldbronn Analytical Division, Waldbronn, Germany» Dona L. Miller. Worldwide
`Customer Loveland, Division. Mountain View. California» Michael P. Moore, VXI Systems Division, Loveland, Colorado» Shelley I. Moore, San Diego Printer Division.
`San Diego, Systems • M. Shahid Mujtaba. HP Laboratories, Palo Alto. California » Steven J. Narciso, VXI Systems Division. Loveland. Colorado » Danny J- Oldfield,
`Colorado Springs Division, Colorado Springs, Colorado » Garry Orsolini, Software Technology Division, Roseville, California • Han Tian Phua. Asia Peripherals Division,
`Singapore • Germany Software HP Laboratories, Palo Alto. California » Gunter Riebesell. Boblingen Instruments Division. Boblingen, Germany • Marc Sabatella. Software
`Engineering Stenton, Division, Fort Collins, Colorado» Michael B. Saunders. Integrated Circuit Business Division, Corvallis, Oregon» Philip Stenton, HP Laboratories
`Bristol, Willits, England» Stephen R. Undy, Systems Technology Division, FortCollins, Colorado» Jim Willits, Network and System Management Division, fort Collins,
`Colorado • Oregon Yanagawa, Kobe Instrument Division, Kobe. Japan » Dennis C. York, Corvallis Division, Corvallis. Oregon • Barbara Zimmer. Corporate Engineering.
`Palo Alto, California
`
`©Hewlett-Packard Company 1995 Printed in U.S.A..
`
`April 1995 Hewlett-Packard Journal
`
`© Copr. 1949-1998 Hewlett-Packard Co.
`
`

`

`I Development of a Low-Cost, High-Performance, Multiuser Business Server System, by Dennis
`A. Bowers, Gerard M. Enkerlin, and Karen L Murillo
`
`| R HP Applications Smalltalk: A Tool for Developing Distributed Applications by Eileen Keremitsis
`and Ian J. Fuller
`
`j Object Management Group
`
`j A Software Solution Broker for Technical Consultants, by Manny Yousefi, Adel Ghoneimy, and
`Wulf Hender
`
`| HP Software Solution Broker Accessible Products
`
`J / Bugs Jack Black and White: Imaging 1C Logic Levels with Voltage Contrast, by Jack D. Benzel
`
`1 H"? Component and System Level Design-for-Testability Features Implemented in a Family of
`I U / Workstation Products, by Bulent I. Dervisoglu and Michael Ricchetti
`
`Departments
`
`4 I n t h i s I s s u e
`5 C o v e r
`5 W h a t ' s A h e a d
`1 1 4 A u t h o r s
`
`T h e H e w l e t t - P a c k a r d J o u r n a l i s p u b l i s h e d b i m o n t h l y b y t h e H e w l e t t - P a c k a r d C o m p a n y t o r e c o g n i z e t e c h n i c a l c o n t r i b u t i o n s m a d e b y H e w l e t t - P a c k a r d
`(HP) personnel. While the information found in this publication is believed to be accurate, the Hewlett-Packard Company disclaims all warranties of
`merchantability and fitness for a particular purpose and all obligations and liabilities for damages, including but not limited to indirect, special, or
`consequential damages, attorney's and expert's fees, and court costs, arising out of or in connection with this publication.
`
`Subscriptions: The Hewlett-Packard Journal is distributed free of charge to HP research, design and manufacturing engineering personnel, as well as to
`qualified address individuals, libraries, and educational institutions. Please address subscription or change of address requests on printed letterhead (or
`include the submitting card) to the HP headquarters office in your country or to the HP address on the back cover. When submitting a change of address,
`please not your zip or postal code and a copy of your old label. Free subscriptions may not be available in all countries.
`
`The Hewlett-Packard Journal is available online via the World-Wide Web (WWW) and can be viewed and printed with Mosaic. The uniform resource
`locator (URL) for the Hewlett-Packard Journal is http://www.hp.com/hpj/Journal.html.
`
`S u b m i s s i o n s : w i t h a r t i c l e s i n t h e H e w l e t t - P a c k a r d J o u r n a l a r e p r i m a r i l y a u t h o r e d b y H P e m p l o y e e s , a r t i c l e s f r o m n o n - H P a u t h o r s d e a l i n g w i t h
`HP-related contact or solutions to technical problems made possible by using HP equipment are also considered for publication. Please contact the
`E d i t o r b e f o r e a r t i c l e s s u c h a r t i c l e s . A l s o , t h e H e w l e t t - P a c k a r d J o u r n a l e n c o u r a g e s t e c h n i c a l d i s c u s s i o n s o f t h e t o p i c s p r e s e n t e d i n r e c e n t a r t i c l e s
`a n d m a y a r e l e t t e r s e x p e c t e d t o b e o f i n t e r e s t t o r e a d e r s . L e t t e r s s h o u l d b e b r i e f , a n d a r e s u b j e c t t o e d i t i n g b y H P
`
`Copyright fv granted Hewlett-Packard Company. All rights reserved. Permission to copy without fee all or part of this publication is hereby granted provided
`that 1) advantage; Company are not made, used, displayed, or distributed for commercial advantage; 2) the Hewlett-Packard Company copyright notice and the title
`of the stating and date appear on the copies; and 3) a notice appears stating that the copying is by permission of the Hewlett-Packard Company.
`
`Please Journal, inquiries, submissions, and requests to: Editor, Hewlett-Packard Journal, 3000 Hanover Street, Palo Alto, CA 94304 U.S.A.
`
`© Copr. 1949-1998 Hewlett-Packard Co.
`
`April 1995 Hewlett-Packard Journal 3
`
`

`

`Real-Time Software MPEG Video
`Decoder on Multimedia-Enhanced PA
`7100LC Processors
`With including combination of software and hardware optimizations, including the
`availability of PA-RISC multimedia instructions, a software video player
`running on a low-end workstation is able to play MPEG compressed video
`at 30 frames/s.
`
`by Ruby B. Lee, John P. Beck, Joel Lamb, and Kenneth E. Severson
`
`Traditionally, computers have improved productivity by
`helping people compute faster and more accurately. Today,
`computers can further improve productivity by helping
`people communicate better and more naturally. Towards
`this end, at Hewlett-Packard we have looked for more natu
`ral ways to integrate communication power into our desktop
`machines, which would allow a user to access distributed
`information more easily and communicate with other users
`more readily.
`
`We felt that adding audio, images, and video information
`would enrich the information media of text and graphics
`normally available on desktop computers such as work
`stations and personal computers. However, for such en
`riched multimedia communications to be useful, it must be
`fully integrated into the user's normal working environment.
`Hence, as the technology matured we decided to integrate
`increasing levels of multimedia support into both the user
`interface and the basic hardware platform.
`
`In terms of user interface, we integrated a panel of multi
`media icons hito the HP VUE standard graphical user inter
`face, which comes with all HP workstations. These multi
`media icons are part of the HP MPower product.1 HP
`MPower enables a workstation user to receive and send
`faxes, share printers, access and manipulate images, hear and
`send voice and CD-quality stereo audio, send and receive
`multimedia email, share an X window or an electronic white
`board with other distributed users, and capture and play back
`video sequences. The HP MPower software is based on a
`client/ server model, in which one server can service around
`20 clients, which can be workstations or X terminals.
`
`hi terms of hardware platforms, we integrated successive
`levels of multimedia support into the baseline PA-RISC work
`stations.2-3'4 First, we integrated support for all the popular
`image formats such as JPEG (Joint Photographic Experts
`Group)t compressed images.5 Then, we added hardware
`and software support for audio, starting with 8-kHz voice-
`quality audio, followed by support for numerous audio for
`mats including A-law, (i-law, and 16-bit linear mode, with up
`to 48-kHz mono and stereo. This allowed high-fidelity,
`
`t JPEG is an international digital image compression standard for continuous-tone (multi
`level) still ¡mages (grayscale and color).
`
`44.1-kHz stereo, 16-bit CD-quality audio to be recorded,
`manipulated, and played back on HP workstations. At the
`same time, we supported uncompressed video capture and
`playback.
`
`In January 1994, HP introduced HP MPower 2.0 and the
`entry-level enterprise workstation, the HP 9000 Model 712,
`which is based on the multimedia-enhanced PA-RISC pro
`cessor known as the PA 7100LC.6'7'8 The video player inte
`grated in the MPower 2.0 product is the first product that
`achieves real-time MPEG-1 (Moving Picture Experts
`Group)9 video decompression via software running on a
`general-purpose processor. Typically real-time MPEG-1 de
`compression is achieved via special-purpose chips or
`boards. Previous attempts at software MPEG-1 decompres
`sion did not attain real-time rates.10 The fact that this is
`achieved by the low-end Model 712 workstation is significant.
`
`In this paper, we discuss the support of MPEG-compressed
`video as a new (video) data type. In particular, we discuss
`the technology that enables the video player integrated into
`the HP MPower 2.0 product to play back MPEG-compressed
`video at real-time rates of up to 30 frames per second.
`
`Digital Video Standards
`We decided to focus on the MPEG digital video format be
`cause it is an ISO (International Standards Organization)
`standard, and it gives the highest video fidelity at a given
`compression ratio of any of the formats that we evaluated.
`MPEG also has broad support from the consumer electron
`ics, telecommunications, cable, and computer industries.
`The high compression capability of MPEG translates into
`lower storage costs and less bandwidth needed for transmit
`ting video on the network. These characteristics make
`MPEG an ideal format for addressing the need for detail in
`the video used in technical workstation markets and com
`puter-based training in commercial workstation markets.
`
`MPEG is one of several algorithmically related standards
`shown in Fig. 1. All of these digital video compression stan
`dards use the discrete cosine transform (DCT) as a funda
`mental component of the algorithm. Alternatives to discrete
`cosine-based algorithms that we looked at include vector
`quantization, fractals, and wavelets. Vector quantization
`
`60 April 1995 Hewlett-Packard Journal
`
`© Copr. 1949-1998 Hewlett-Packard Co.
`
`

`

`(image) of an H.261 sequence is for all practical purposes a
`highly compressed lossy JPEG image. Subsequent frames
`are built from image fragments (blocks) that are either
`JPEG-like or are differences from the image fragments in
`pre\ious frames. Most video sequences have high frame-to-
`frame coherence. This is especially true for video conferenc
`ing. Because the encoding of the movement of a piece of an
`image requires less data than an equivalent JPEG fragment.
`H.261 achieves higher isual fidelity for a given bandwidth
`than does motion JPEG. Since the encoding of the differ
`ences is always based on the previous frames, the technique
`is called/o ncard differencing.
`
`The MPEG-1 specification goes even further than H.261 in
`allowing sophisticated techniques to achieve high fidelity
`with fewer bits. In addition to forward differencing, MPEG-1
`allows backward differencing (which relies on information
`in a future frame) and averaging of image fragments. (For
`ward and backward differencing are described in more de
`tail in the next section.) MPEG-1 achieves quality compara
`ble to a professionally reproduced VHS videotape even at a
`single-speed CD-ROM data rate (1.5 Mbits/s).9'11 MPEG-1
`also specifies encodings for high-fidelity audio synchro
`nized with the video.
`
`MPEG-2 contains additional specifications and is a superset
`of MPEG-1. The new features in MPEG-2 are targeted at
`broadcast television requirements, such as support for
`frame interleaving similar to analog broadcast techniques.
`With widespread deployment of MPEG-2, the digital revolu
`tion for video may be comparable to the digital audio revolu
`tion of the last decade.
`
`The approximate bandwidths required to achieve a level of
`subjective visual fidelity for motion JPEG, H.261, MPEG-1,
`and MPEG-2 are shown in Fig. 2. Motion JPEG will primarily
`be used for cases in which accurate frame editing is impor
`tant such as video editing. H.261 will be used primarily for
`video conferencing, but it also has potential for use in video
`mail. MPEG-1 and MPEG-2 will be used for publishing,
`where fidelity expectations have been set by consumer ana
`log video tapes, computer-based training, games, movies on
`CD, and video on demand.
`
`MPEG Compression
`MPEG has two classes of frames: intracoded and non-
`inlracoded frames (see Fig. 3). Intracoded frames, also called
`I-frames, are compressed by reducing spatial redundancy
`within the frame itself. I-frames do not depend on compari
`sons with past or "reference" frames. They use JPEG-type
`compression for still images.5
`
`— International
`Image Standard
`
`Fig. transform. Digital video standards based on the discrete cosine transform.
`
`algorithms are popular on older computer architectures be
`cause they require less computing power to decompress, but
`this advantage is offset by poorer image quality at low band
`width (high compression) compared to MPEG for practical
`vector quantization methods. Algorithms based on wavelet
`and fractal technology have the potential to deliver video
`fidelity comparable to MPEG, but there is presently a lack of
`industry consensus on standardization, a key requirement for
`our use.
`Another advantage of a high-performance implementation of
`MPEG is the ability to leverage the improvements to the
`other DCT-based algorithms. Although the relationships
`shown in Fig. 1 do not represent a true hierarchy of algo
`rithms is useful for illustrating increased complexity as one
`moves from JPEG to MPEG-2, or from H.261 to MPEG-2.
`
`All of these formats have much in common, such as the use
`of the DCT for encoding. The visual fidelity of the algorithms
`was the key selection criterion and not ease of implementa
`tion or performance on existing hardware.
`
`Although JPEG supports both lossy and lossless compres
`sion, the term JPEG is typically associated with the lossy
`specification, t The primary goal of JPEG is to achieve high
`compression of photographic images with little perceived
`loss of image fidelity. Although it is not an ISO standard, by
`convention, a sequence of JPEG lossy images to create a
`digital video sequence is called motion JPEG, or MJPEG.
`
`H.261 is a digital video standard from the telecommunica
`tions standards body ITU-TSS (formerly known as CCITT).
`H.261 is one of a suite of conferencing standards that make
`up the umbrella H.320 specification. H.261 is often referred
`to as P*64 (where P is an integer) because it was designed
`to fit into multiples of 64 kbits/s bandwidth. The first frame
`t In lossless compression, decompressed data is identical to the original image data. In lossy
`compression, decompressed data is a good approximation of the original image data.
`
`30,000
`
`- 1 0 , 0 0 0
`
`! 3 , 0 0 0 -
`
`: 1 , 0 0 0
`
`Ã(cid:173) - 100
`
`•
`
`Uses
`
`ction and Broadcast
`^ 0 ^ Â ¡ 7 . 2 | P r o d u
`
`MPEG-1
`
`I Computer Based Training,
`f Analysis, and Monitoring
`
`} Conferencing and Video Mail
`
`Fig. 2. Compressed video band-
`uiillh versus subjective visual
`Fidelity. The ideal format achieves
`exceptional visual fidelity at the
`lowest bandwidth.
`
`30
`10
`
`P o o r A d e q u a t e G o o d V e r y G o o d E x c e l l e n t E x c e p t i o n a l
`Visual Fidelity
`
`© Copr. 1949-1998 Hewlett-Packard Co.
`
`April 1995 Hewlett-Packard Journal 61
`
`

`

`Forward Prediction
`
`Bidirectional Prediction
`
`| I = lntracoded Frame (Like JPEG)
`
`I | P = Predicted Frame
`
`J B = B¡d¡rectionally Predicted Frame
`
`Nonintracoded frames are further divided into P-frames and
`B-frames. P-frames are predicted frames based on compari
`sons with an earlier reference frame (an intracoded or pre
`dicted frame). By considering temporal redundancy in addi
`tion to spatial redundancy, P-frames can be encoded with
`fewer bits. B-frames are bidirectionally predicted frames
`that require one backward reference frame and one forward
`reference frame for prediction. A reference frame can be an
`I-frame or a P-frame, but not a B-frame. By detecting the
`motion of blocks from both a frame that occurred earlier
`and a frame that will be played back later in the video
`sequence, B-frames can be encoded in fewer bits than I- or
`P-frames.
`
`Each frame is divided into macroblocks of 16 by 16 pixels
`for the purposes of motion estimationt in MPEG compression
`and motion compensation in MPEG decompression. A frame
`with only I-blocks is an I-frame, whereas a P-frame has P-
`blocks or I-blocks, and a B-frame has B-blocks, P-blocks, or
`I-blocks. For each P-block in the current frame, the block in
`the reference frame that matches it best is identified by a
`motion vector. Then the differences between the pixel values
`in the matching block in the reference frame and the current
`block in the current frame are encoded by a discrete cosine
`transform.
`
`The color space used is the YCbCr color representation
`rather than the RGB color space, where Y represents the
`luminance (or brightness) component, and Cb and Cr repre
`sent the chrominance (or color) components. Because
`human perception is more sensitive to luminance than to
`chrominance, the Cb and Cr components can be subsampled
`in both the x and y dimensions. This means that there is one
`Cb value and one Cr value for every four Y values. Hence, a
`16-by-16 macroblock contains four 8-by-8 blocks of Y, and
`only one 8-by-8 block of Cb and one 8-by-8 block of Cr val
`ues (see Fig. 4). This is a reduction from the twelve 8-by-8
`blocks (four for each of the three color components) if Cb
`
`Fig. 3. MPEG frame sequencing.
`
`and Cr were not subsampled. The six 8-by-8 blocks in each
`16-by-16 macroblock then undergo transform coding.
`
`Transform coding concentrates energy in the lower fre
`quencies. The transformed data values are then quantized by
`dividing by the corresponding quantization coefficient. This
`results in discarding some of the high-frequency values, or
`lower-frequency but low-energy values, since these become
`zeros. Both transform coding and quantization enable further
`compression by run-length encoding of zero values.
`
`Finally, the nonzero coefficients of an 8-by-8 block used in
`the discrete cosine transform can be encoded via variable-
`length entropy encoding such as Huffman coding. Entropy
`encoding basically removes coding redundancy by assigning
`the code words with the fewest number of bits to those co
`efficients that occur most frequently.
`
`X = Luminance (Y|
`0 = Chrominance (Cb, Cr)
`
`t Motion estimation uses temporal redundancy to estimate the movement of a block from
`one frame to the next.
`
`Fig. with Subsampling of the chrominance components (Cb, Cr) with
`respect to the luminance (Y) component.
`
`62 April 1995 Hewlett-Packard Journal
`
`© Copr. 1949-1998 Hewlett-Packard Co.
`
`

`

`we found ways to reduce the execution time of the most
`frequent operation sequences. The application of algorithm
`enhancements, software tuning, and projected hardware
`enhancements was iterated until we attained our goal of
`being able to decompress at a rate greater than 15 frames/s
`via software.
`
`Algorithm and Software Optimizations
`In terms of MPEG \ideo algorithms, we improved on the
`Huffman decoder, the motion compensation, and the inverse
`discrete cosine transform. A faster Huffman decoder based
`on a hybrid of table lookup and tree-based decoding is used.
`The lookup table sizes were chosen to reduce cache misses.
`For motion compensation, we sped up the pixel averaging
`operations.
`
`For the inverse discrete cosine transform, we use a faster
`Fourier transform, which significantly reduces the number
`of multiplies for each two-dimensional 8-by-8 inverse dis
`crete cosine transform. In addition, we use the fact that the
`8-by-8 inverse transform matrices are frequently sparse to
`further reduce the multiplies and other operations required.
`
`The MPEG audio decompression is also done in software.
`This algorithm was improved by using a 32-point discrete
`cosine transform to speed up the subband filtering.12
`
`In terms of software tuning, we "flattened" the code to re
`duce the number of procedure calls and returns, and the
`frequent building up and tearing down of contexts present in
`the original MPEG code. We also did "strength reductions"
`like as multiplications to simpler operations such as
`shift and add or table lookup.
`
`The last column of Table I shows the percentage of execu
`tion time spent in each of the six MPEG decompression steps
`after the algorithm and software tuning improvements were
`made. The first two columns of Table I show the millions of
`instructions executed in each of the six decompression steps
`and the percent of the total instructions executed (path
`length) each step represents. The input video sequence was
`an MPEG-compressed clip of a football game. The total time
`taken was 7.45 seconds on an HP 9000 Model 735 99-MHz
`PA-RISC workstation, with 256K bytes of instruction cache
`and 256K bytes of data cache.
`
`Table I
`I n s t r u c t i o n s a n d T i m e S p e n t i n e a c h M P E G D e c o m p r e s s i o n
`Step on an HP 9000 Model 735
`
`M i l l i o n s o f P a t h L e n g t h T i m e ( % )
`I n s t r u c t i o n s ( % )
`
`H e a d e r d e c o d e 0 . 6 0 . 1 0 . 1
`H u f f m a n d e c o d e 5 5 . 3 1 0 . 2 7 . 5
`I n v e r s e q u a n t i z a t i o n 8 . 7 1 . 6 2 . 4
`I n v e r s e d i s c r e t e 2 0 6 . 5 3 8 . 3 3 8 . 7
`
`MPEG Decompression
`MPEG decompression reverses the functional steps taken
`for MPEG compression. There are six basic steps involved
`in MPEG decompression.
`1. The MPEG header is decoded. This gives information
`such as picture rate, bit rate, and image size.
`
`2. The video data stream is Huffman or entropy decoded
`from variable-length codes into fixed-length numbers. This
`step includes run-length decoding of zeros.
`
`3. Inverse quantization is performed on the numbers to
`restore them to their original range.
`
`4. An inverse discrete cosine transform is performed on the
`8-by-8 blocks in each frame. This converts from the frequency
`domain back to the original spatial domain. This gives the
`actual pixel values for I-blocks, but only the differences for
`each pixel for P-blocks and B-blocks.
`
`5. Motion compensation is performed for P-blocks and B-
`blocks. The differences calculated in step 4 are added to the
`pixels in the reference block as determined by the motion
`vector for P-blocks and to the average of the forward and
`backward reference blocks for B-blocks.
`
`6. The picture is displayed by doing a color conversion from
`YCbCr coordinates to RGB color coordinates and writing to
`the frame buffer.
`
`Methodology
`Our philosophy was to improve the algorithms and tune the
`software first, resorting to hardware support only if neces
`sary. We set a goal of 10 to 15 frames/s for software MPEG
`video decompression because this is the rate at which mo
`tion appears smooth rather than jerky.
`
`We started by measuring the performance of the MPEG soft
`ware we had purchased. This software initially took two
`seconds to decode one frame (0.5 frame/s) on an older
`50-MHz Model 720 workstation. This decoding was for video
`only and did not include audio. Profiling indicated that the
`inverse discrete cosine transform (step 4) took the largest
`chunk of the execution time, followed by display (step 6),
`followed by motion compensation (step 5). The decoding of
`the MPEG headers was insignificant.
`
`With this data we set out to optimize every step in the MPEG
`decompression software. After we applied all the algorithm
`enhancements and software tuning, we measured the MPEG
`decode software again. While we had achieved an order of
`magnitude improvement, the rate of 4 to 5 frames/s was not
`sufficient to meet our goal.
`
`Hence, we looked at possible multimedia enhancements to
`the basic PA-RISC processor and other system-level en
`hancements that would not only speed up MPEG decoding,
`but also be generally useful for improving performance in
`other computations. In addition, any chip enhancements we
`added could not adversely impact the design schedule, com
`plexity, cycle time, and chip size of the PA-RISC processor
`we were targeting, the PA 7100LC, which was already deep
`into its implementation phase at the time. The PA 7100LC is
`described in detail in the article on page 12.
`
`We approached this problem by studying the distribution of
`operations executed by the software MPEG decoder. Then,
`
`© Copr. 1949-1998 Hewlett-Packard Co.
`
`April 1995 Hewlett-Packard Journal 63
`
`

`

`The largest slice of execution time (38.7%) and the largest
`chunk of instructions executed (38.3%) were still the inverse
`discrete cosine transform. We studied the frequencies of
`generic operations in this group and attempted to execute
`them faster. This resulted in new PA-RISC processor in
`structions for accelerating multimedia software.
`
`PA-RISC Processor Enhancements
`The new processor multimedia instructions implemented
`in the PA 7100LC processor allow simple arithmetic opera
`tions to be executed in parallel on subword data in the stan
`dard integer data path. In particular, the integer ALU is parti
`tioned so that it can execute a pair of arithmetic operations
`in a single cycle with a single instruction. The arithmetic
`operations accelerated in this way are add, subtract, aver
`age, shift left and add, and shift right and add. The latter two
`operations are effective in implementing multiplication by
`constants.
`
`PA-RISC Multimedia Extensions 1.0. The PA 7100LC PA-RISC
`processor chip contains some instructions that operate inde
`pendently and in parallel on two 16-bit data fields within a
`32-bit register. These operations are independent in that bits
`carried or shifted out of one of the fields never affects the
`result in the other field. These operations occur in parallel in
`that a single instruction computes both 16-bit fields of the
`result. Table II summarizes these instructions.
`
`HADD does two parallel 16-bit additions on the left and the
`right halves of registers ra and rb, placing the two 16-bit re
`sults into the left and right halves of register rt.
`
`HSU B does two parallel 16-bit subtractions on the left and
`right halves of registers ra and rb, placing the two 16-bit re
`sults into the left and right half of register rt.
`
`Both HADD and HSUB perform modulo arithmetic (modulus
`216), that is, the result wraps around from the largest number
`back to the smallest number and vice versa. This is the usual
`mode of operation of twos complement adders when over
`flow is ignored.
`
`HADD and HSUB also have two saturation arithmetic options.
`With the signed saturation option, HADD.ss, both operands
`and the result are considered signed 16-bit integers. If the
`result cannot be represented as a signed 16-bit integer, it is
`clipped to the largest positive value (215-1) if positive over
`flow occurs, or it is clipped to the smallest negative value
`(-215) if negative overflow occurs.
`
`With the unsigned saturation option, HADD.us, the first oper
`and (ra) is considered an unsigned 16-bit integer, the second
`operand (rb) is considered a signed 16-bit integer, and the
`result (in rt) is considered an unsigned 16-bit integer. If the
`result cannot be represented as an unsigned 16-bit integer, it
`is clipped to the largest unsigned value (216-1) if positive
`overflow occurs, or it is clipped to the smallest unsigned
`value (0) if negative overflow occurs.
`
`The signed saturation and unsigned saturation options for
`parallel halfword subtraction are defined similarly.
`
`HAVE, or halfword average, gives the average of each pair of
`halfwords in ra and rb. It takes the sum of parallel halfwords
`and does a right shift of one bit before storing each 16-bit
`result into rt. During the one-bit right shift, the carry is
`
`Table II
`PA-RISC Multimedia Instructions in PA 7100LC
`ra contains a1; a2
`rb contains b1; b2
`rt contains t1; t2
`Parallel Operation
`
`Instruction
`
`HADD ra,rb,rt
`
`HADD.ss ra,rb,rt
`
`HADD.us ra,rb,rt
`
`HSUB ra.rb.rt
`
`HSUB.ssra,rb,rt
`
`HSUB. us ra.rb.rt
`
`tl =(a1+b1)mod216;
`I2 = (a2+b2)mod216;
`
`tl = IF(a1+b1) > (215-1)THEI\I(215-1)
`ELSEIF(a1+b1) < -215THEN (-215)
`ELSE(a1+b1);
`t2=IF(a2+b2) > (215-1)THEN (215-1)
`ELSEIF(a2+b2) < -215THEN (-215)
`ELSE(a2+b2);
`
`t1=IF(a1+b1) > (216-1)THEN(216-1)
`ELSEIF(a1+b1) < OTHENO
`E L S E ( a H b l ) ;
`t2=IF(a2+b2) > (216-1)THEN (216-1)
`ELSEIF(a2+b2) < OTHENO
`ELSE (a2+b2);
`
`tl =(a1-b1)mod216;
`t2 = (a2-b2)mod216;
`
`tl =IF(a1-b1) > (215-1)THEN(215-1)
`ELSEIF(al-bl) < -215THEN (-215)
`E L S E ( a l - b l ) ;
`t2=IF(a2-b2) > (215-1)THEN(215-1)
`ELSEIF(a2-b2) < -215THEN (-215)
`ELSE(a2-b2);
`
`tl =IF(al-b1) > (216-1)THEN(216-1)
`E L S E I F ( a l - b l ) < O T H E N O
`E L S E ( a l - b l ) ;
`t2=IF(a2-b2) > (216-1)THEN (216-1)
`ELSEIF(a2-b2) < OTHENO
`ELSE (a2-b2);
`
`HAVE ra.rb.rt
`
`tl =(a1+b1)/2;
`12 = (a2+b2)/2;
`
`HSLkADD ra.k.rb.rt
`
`tl =
`
`(for k = 1,2, or 3)
`
`tl =(a1>k) + b1;
`t2 = (a2>k) + b2;
`(for k =1,2, or 3)
`
`HSRkADD ra.k.rb.rt
`
`ss = signed saturation option
`us = unsigned saturation
`
`shifted in on the left and unbiased rounding is performed on
`the least-significant bit on the right. Because the carry is
`shifted in, no overflow can occur in the HAVE instruction.
`
`HSLkADD, or halfword shift left and add, allows one operand
`to be shifted left by k bits (where k is 1, 2, or 3) before being
`added to the other operand.
`
`HSRkADD, or halfword shift right and add, allows one operand
`to be shifted right by k bits (wher