Model-Based Design for Embedded Systems- P24

Chia sẻ: Cong Thanh | Ngày: | Loại File: PDF | Số trang:30

Thêm vào BST

Báo xấu

57
lượt xem 4
download

Download Vui lòng tải xuống để xem tài liệu đầy đủ

Model-Based Design for Embedded Systems- P24:The unparalleled flexibility of computation has been a key driver and feature bonanza in the development of a wide range of products across a broad and diverse spectrum of applications such as in the automotive aerospace, health care, consumer electronics, etc.

Chủ đề:

Bình luận(0) Đăng nhập để gửi bình luận!

Lưu

Nội dung Text: Model-Based Design for Embedded Systems- P24

666 Model-Based Design for Embedded Systems Fourier transform can be implemented by one of the numerous fast Fourier transform (FFT) techniques. The computational order of the FFT for a 2D input is O(N2 log2 N), obviously more efficient when compared to the direct integration method. We show this speed increase later through an example. In continuous theory, the angular spectrum method is an exact solution of the Rayleigh–Sommerfeld formulation. However, when solving the algo- rithm on a digital computer, a discrete Fourier transform (DFT) must be used, resulting in the accuracy of the angular spectrum method being dependent on the resolution, or spacing, of the aperture and observation plane meshing. We call the physical size of the aperture and observation planes the “bound- ing box,” defining the size of the optical wave front being propagated. Since the complex wave function is only nonzero for a finite space in the bounding box, the signal is not always bandwidth limited, and the Nyquist sampling theory does not always apply. It can be shown, however, that the resolu- tion of the aperture and observation meshing must be λ/2 or smaller [39]. For many simulation systems without large degrees of tilt and hard diffrac- tive apertures, the resolution can be coarser. In systems with high tilts, the resolution is most sensitive. With a mesh spacing of λ/2, the angular spec- trum decomposition will model plane waves propagating from the aperture to the observation plane in a complete half circle, that is, between –90 and +90 degrees. Other inaccuracies that can occur when using a DFT are aliasing and win- dow truncation. Aliasing occurs when frequencies exist greater than the criti- cal sampling frequency. In this case, these high frequencies are “folded over” into the sampled frequency range [40]. The effect of this is seen in our simu- lations as optical power “reflecting” off of the walls of the bounding box. If significant optical power reflects off the wall, interference between the prop- agating beam and these reflections can occur, resulting in inaccurate optical waveforms. The same effect can be seen when the bounding box truncates the signal. Truncation occurs when the waveform propagates into the edges of the bounding box. The simplest solution to ensure accurate results is having sufficient zero padding around the optical waveform, reducing the chance the waveform is aliased or truncated by walls of the bounding box. In Chatoyant, the user can choose between using the Gaussian or scalar diffractive (angular spectrum) methods during simulation. The components in the optical library support both representations in the optical signal message class. Using these models we can simulate and analyze a variety of heterogeneous systems as presented in the next section. 20.2.7 Simulations and Analysis of Optical MEM Systems In this section, we show how Chatoyant can model and simulate complete mixed-signal systems. The first system uses both electrical and optical sig- nals to simulate a complete “4f” optoelectronic link which uses a four focal length image relaying optical system. The second example, building from the
CAD Tools for Multi-Domain Systems on Chips 667 two signal 4f link, adds mechanical signals for simulation and analysis of an optical MEM system. This set of example systems is centered on an optical MEM scanning mirror. With this device we are able to simulate an optical scanning system and a self-aligning optical detection system. These systems show the ability to model a mixed system of mechanical MEMs, optics, and electronic feedback. The last example shows the power of the angular spec- trum technique to model diffractive optical systems with the speed and accu- racy required to perform system-level design. 20.2.7.1 Full Link Example A complete optoelectronic simulation of a 4f optical communication link in Chatoyant is presented in Figure 20.12. The distance between the vertical cavity surface emitting laser (VCSEL) array and the first lens and the dis- tance between the second lens and the detector array are both 1 mm. The distance between the lenses is 2 mm, with both lenses having a focal length of 1 mm, giving a 4f system. The top third of the figure shows the system as represented in Chatoyant. Each icon represents a component model, and each line represents a signal path (either optical or electrical) connecting the outputs of one component to the inputs of the next. Several of the icons, such as the VCSELs and receivers, model the optoelectronic components them- selves, while others, such as the output graph, are used to monitor and dis- play the behavior of the system. The input to the system is an electrical signal with speed varying from 300 MHz to 1.5 GHz. A Gaussian noise with vari- ance of 0.5 V has been added to the multistage driver system to show the ability of our models to respond to arbitrary waveforms. In the center of the figure, three snapshots (before the VCSEL, after the VCSEL, and after the detector) show the behavior of the CMOS drivers under Driver Gaussian waist analysis Digital PGM + Power analysis VCSEL Receiver 4f optical system FIGURE 20.12 Chatoyant analysis of optoelectronic 4f communications link.
668 Model-Based Design for Embedded Systems a 300 MHz noisy signal. In these snapshots, one can see the amplification of the system noise through the CMOS drivers, the clipping of subthreshold noise in the VCSEL, and the frequency response on the quality of the received signal. This last observation is better seen in the three eye diagrams, shown at the bottom of Figure 20.12, analyzed at 300 MHz, 900 MHz, and 1.5 GHz. For the component values chosen, the system operates with reasonable BER up to about 1 GHz. For this 4f system, the VCSEL and driver circuits explicitly model the effects of bias current and temperature on the optoelectric conversion, L-I efficiency, of the lasers. Figure 20.13 shows the effects of temperature, T, and current bias, Ib , on the bit error rate (BER) of the link. Generally, the fre- quency response of the link is dominated by the design of the receiver circuit; however it is interesting to note that both the VCSEL temperature and bias have a significant effect on system performance, because of their impact in the power through the link. Perhaps most interesting is the fact that increas- ing bias current does not always correspond to better performance over the whole range of frequencies examined. Note that the curve for 1 mA bias offers the best performance below 600 MHz; however, the 0.5 mA bias (the nominal threshold of the VCSEL) crosses the curve for 1 mA and achieves the best performance at higher frequencies. As an example of mechanical tolerancing, we analyze the system with varying-sized photodetectors (50, 30, and 20 μm). The detectors are displaced from +10 μm to +100 μm in detector position along the axis of optical prop- agation. This results in defocusing of the beam relative to the detector array. We calculate both the insertion loss and the worst case optical crosstalk as the detectors are displaced. The results are shown in Figure 20.14. Systems can be further analyzed for their sensitivity to mechanical tolerances using a Monte Carlo tolerancing method described in [8,9]. Two additional analyses are also shown in the Chatoyant representation in Figure 20.12. The first is the beam profile analysis, which graphically dis- plays one beam’s waist as it propagates between components, showing the possibility of clipping at the lenses. The second analysis shows the optical signals as they strike the detector array. This analysis also gives the user the amount of optical power captured on each of the detectors. From this analy- sis, optical crosstalk and system insertion loss can be calculated. 20.2.7.2 Optical Beam Steering/Alignment System A torsion-scanning mirror is a micromachined 2D mirror built upon a micro-elevator by self assembly (MESA) structure [41,42]. The mirror and MESA structures are shown in Figure 20.15a and b, respectively. The scan- ning mirror can tilt along the torsion bars in both the x and y directions and is controlled electrostatically through four electrodes beneath the mirror, outlined in Figure 20.15a by the dashed boxes. For example, the mirror tilts in the positive x direction when voltage is applied to electrodes 1 and 2, and the
CAD Tools for Multi-Domain Systems on Chips 669 BER vs. frequency at various current bias 1.E – 20 BER vs. frequency at VCSEL temperatures 1.E – 16 1.E – 20 1.E – 12 BER 1.E – 16 1.E – 08 1.E – 12 BER 1.E – 04 1.E – 08 1.E + 00 100 300 500 700 900 1100 1300 1500 1.E – 04 Frequency (MHz) 1.E + 00 100 300 500 700 900 1100 1300 1500 BER (lb = 0.1 mA) BER (lb = 0.25 mA) Frequency (MHz) BER (lb = 1.0 mA) BER (lb = 1.5 mA) BER (T = 40 C) BER (T = 70 C) BER (T = 100 C) BER (lb = 0.5 mA) FIGURE 20.13 BER versus frequency at different VCEL temperatures and current biases. Crosstalk vs. detector displacement Insertion loss vs. detector displacement 10 30 50 70 10 30 50 70 0 Insertion loss (db) 0 Crosstalk (dB) –3 –25 –6 –50 –9 –12 –75 –15 –100 –18 +μm displaced in optical axis ±μm displaced in optical axis 50 um Det 30 um Det 20 um Det 50 um Det 30 um Det 20 um Det FIGURE 20.14 Insertion and crosstalk versus mechanical tolerancing. (From Kurzweg, T.P. et al., J. Model. Simul. Micro-Syst., 2, 21, 2001. With permission.) mirror tilts in the negative y direction when voltage is applied to electrodes 1 and 4. The MESA structure is shown in Figure 20.15b. The mirror is elevated by four scratch drive actuator (SDA) sets pushing the support plates together, allowing for the scanning mirror to buckle and rise up off the substrate [43]. The MESA structure’s height is required to be large enough such that the tilt of the mirror will not cause the mirror to hit the substrate. Post fabrication system alignment can also be performed by the MESA structure. Figure 20.16 shows a drawing of the torsion-scanning mirror system. On the left one can see one VCSEL emitting light vertically through a lenslet, and a prism that reflects off a plane mirror. The light is then reflected off of the optical MEM scanning mirror, back to the plane mirror, and captured through a lenslet and prism onto detectors on the right. With the flexibility of the scanning mirror, this system could act as a switch, an optical scanner,
670 Model-Based Design for Embedded Systems y 3 2 x 4 1 (a) (b) FIGURE 20.15 (a) Scanning torsion mirror, (b) MESA structure. (From Kurzweg, T.P. et al., CAD for optical MEMS, Proceedings of the 36th IEEE/ACM Design Automation Conference (DAC’99), New Orleans, LA, June 20–25, 1999. With permission.) FIGURE 20.16 Scanning mirror system. (From Kurzweg, T.P. et al., CAD for optical MEMS, Proceedings of the 36th IEEE/ACM Design Automation Conference (DAC’99), New Orleans, LA, June 20–25, 1999. With permission.) or a reconfigurable optical interconnect. We have simulated systems using this scanning mirror configuration for switching and self-alignment through optical feedback. We first demonstrate an optical scanning system. In this scanning system, we simulate a single source beam propagating through the 3 × 3 subsystem seen in Figure 20.16. With the appropriate volt- age levels applied to the four electrodes, the scanning mirror tilts and directs the source to any of the nine detectors. This system, as represented in Chatoy- ant, is shown in Figure 20.17. The SDA arrays move the mirror to the correct height for alignment. We control the electrodes with a waveform generator, which applies the appropriate voltages on the four electrodes for the beam to scan or switch in a desired pattern. As an example, we are able to scan a diamond pattern with the wave- forms shown in Figure 20.18. The desired pattern is shown by the white arrow trace on the first output image. The other nine images show snapshots of the detector plane as the diamond pattern is scanned. Dashed lettered lines correspond to time intervals in the waveforms and in the snapshots. Mechanical alignment is critical in this system. For example, the lenslets in this simulation are only 100 μm in diameter. Therefore, when steering the
UCSEL Prism Mirror Mirror Prism Powergrid Const SDA FIGURE 20.17 Scanning system as represented in Chatoyant. (From Kurzweg, T.P. et al., CAD for optical MEMS, Proceedings of the 36th IEEE/ACM Design Automation Conference (DAC’99), New Orleans, LA, June 20–25, 1999. With permission.) CAD Tools for Multi-Domain Systems on Chips 671
672 Model-Based Design for Embedded Systems Electrode 3 Electrode 2 Electrode 4 Electrode 1 A B C D E A B C D E xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: FIGURE 20.18 Scanning waveforms and scanned diamond pattern. (From Kurzweg, T.P. et al., CAD for optical MEMS, Proceedings of the 36th IEEE/ACM Design Automation Conference (DAC’99), New Orleans, LA, June 20–25, 1999. With permission.) beam, precision in the voltage waveforms is needed so that the light, bend- ing through the prism, hits the desired detector’s lenslet. We next simulate a self-aligning system using optical feedback, using the same system setup as seen in Figure 20.16. Such a system could be used as a noise suppression system. The scanning mirror is used to actively align the system, with the electrodes now being controlled by a waveform generator with a programmed control algorithm. The waveform generator receives the power values detected on each of the detectors, determines where the beam is, and which electrodes to apply voltage to in order to steer the beam onto the center detector. The system is considered aligned when the power detected on the center detector matches a threshold value set by the user. The user also specifies, in the control algorithm, the size of the voltage step that will be placed on the corresponding electrodes. With active feedback, the system will keep step- ping enough voltage to the electrodes until the beam is steered onto the cen- ter detector and the system is aligned. The system, as displayed in Chatoyant, is shown in Figure 20.19.
Const FIGURE 20.19 Self-aligning system using optical feedback. (From Kurzweg, T.P. et al., J. Model. Simul. Micro-Syst., 2, 21, 2001. With permis- sion.) CAD Tools for Multi-Domain Systems on Chips 673
674 Model-Based Design for Embedded Systems Time xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: (a) Time xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: (b) Time xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: xv 3.10a: (c) FIGURE 20.20 Self-alignment results. (From Kurzweg, T.P. et al., J. Model. Simul. Micro-Syst., 2, 21, 2001. With permission.) To simulate this self-aligning system, we introduced random offsets in the lenses and in the VCSEL position and observe as the beam moves toward focus on the center detector. Snapshots of the image at the detectors are given in Figure 20.20 for three cases. The first results, shown in Figure 20.20a, are when the second lens is offset 35 μm in the x-direction. Figure 20.20b shows the results of the second lenslet offset in both the −x- and y-direction by 35 μm. The final case has both lenses offset. The first is offset by 5 μm in the x-direction, and the second lens is offset by 35 μm in the −x-direction and 5 μm in the y-direction. The results are seen in Figure 20.20c. Notice that the beam on the final images is not exactly in the center of the middle detector. This is because of the power being detected at this point exceeding the power threshold (98.6%) we set for alignment. 20.2.7.3 Angular Spectrum Optical Simulation of the Grating Light Valve In this section, we simulate and analyze a grating light valve (GLV) sys- tem in Chatoyant. This device has many display applications, including digital projection, HDTV, and vehicle displays. The GLV is simply a MEM
CAD Tools for Multi-Domain Systems on Chips 675 (micro-electrical-mechanical) phase grating made from parallel rows of reflective ribbons. When all the ribbons are in the same plane, incident light that strikes normal to the surface reflects 180 degrees off the GLV. However, if alternating ribbons are moved down a quarter of a wavelength, a “square- well” diffraction pattern is created, and the light is reflected at an angle from that of the incident light. The angle of reflection depends on the width of the ribbons and the wavelength of the incident light. Figure 20.21 shows the rib- bons, from both a top and side view, and also the reflection patterns for both positions of the ribbons. The GLV component is fabricated using standard silicon VLSI technol- ogy, with ribbon dimensions approximately 3–5 μm wide and 20–100 μm long [44]. Each ribbon moves through electrostatic attraction between the ribbon and an electrode fabricated underneath the ribbon. This electrostatic attraction moves the ribbons only a few hundred nanometers, resulting in an approximate switching time of 20 ns. Since the GLV depends on a diffrac- tive phenomenon to direct the light beam, a rigorous modeling technique is required for modeling the GLV system. For the simulation of the GLV, we examine one optical pixel. A projected pixel is diffracted from a GLV composed of four ribbons, two stationary and two that are movable [44]. Each ribbon has a length of 20 μm and a width of 5 μm. Ideally, there is no gap between the ribbons, however, in reality, a gap is present and is a function of the feature size of the fabrication. Although this gap can be modeled in our tool, in these simulations, we provide an ideal GLV simulation with no gap. The GLV is modeled as a phase grating, where the light that strikes the down ribbons propagates a half of a wavelength more than the light that strikes the up ribbons. In our model, light reflecting from the down ribbons is multiplied by a phase term. The phase term is similar to a propagation term through a medium: Udown_ribbon = U exp(j2kd), where d is the distance that the ribbon is moved toward the substrate, typically λ/4 for the GLV. Far-field diffraction theory states that the diffracted angle reflected from the square-well grating is [36]: θ = qλ/a, where q is the diffraction mode Down ribbons Incident Reflected Reflected Incident Reflected Up ribbons 1/4 λ Ribbons (a) (b) (c) FIGURE 20.21 GLV device (a) top view and side view operation for, (b) up ribbons and, (c) down ribbons.
676 Model-Based Design for Embedded Systems (0, +1, +2, +3,. . . ), a is the period of the diffractive grating, and θ is in radians. In the special case of a square well, when light is diffracted by a grating with a displacement of λ/4 (a λ/2 optical path difference after reflection), all the optical power is diffracted from the even modes into the odd modes [45]. In the first simulation, the standard operation of the GLV is verified. We assume an incident plane wave of green light (λgreen 520 nm) striking the grating, with the square-well period defined by the ribbon width, and no gap. We simulate the GLV in both cases, that is, when all the ribbons are on the same plane and when the alternating ribbons are moved downward a distance of λ/4. In this example, the light is reflected off of the grating and propagated 1000 μm to an observation plane. A bounding box of 400 × 400 μm is used, with N equal to 2048. Intensity contours of the observation plane are presented in Figure 20.22a and b. When the grating is moved into the down position, all of the optical power is not transferred into the expected odd far-field diffractive modes. This is seen in the center of Figure 20.22b, as small intensity clusters are scattered between the +1st modes. This scattering is a near-field effect and demonstrates that in this system, light propagating 1000 μm, is not in the far field. If a designer used a tool propagating with the Fraunhofer far-field approximations, these scattering effects would not be detected. For exam- ple, when running the same simulation on LightPipes [46], a CAD tool using the Fraunhofer approximation for optical propagation, only the far-field pat- tern of light diffracted into the 1st and 3rd modes is seen, as presented in Figure 20.22c. When comparing this result to Figure 20.22b, it is shown that far-field approximation is not valid for this propagation distance. Through this example we have shown that using the angular frequency technique, we achieve the full Rayleigh–Sommerfeld accuracy, while obtaining the same computational speed of using the Fraunhofer approximation. To show the advantage of the angular spectrum method, we compare the run time of the above simulation with the run time using the direct inte- gration method. With N = 2048, the FFT simulation takes about 1.5 min. –0.0002 –0.0002 –0.0002 th st 0 mode +1st mode _ +1 mode _ 0.0 0.0 0.0 rd +3rd mode _ +3 mode _ 0.0002 0.0002 0.0002 –0.0002 0.0 0.0002 –0.0002 0.0 0.0002 –0.0002 0.0 0.0002 (a) (b) (c) FIGURE 20.22 GLV operation (a) all ribbons up, (b) alternating ribbons down, (c) Fraun- hofer approximation.
CAD Tools for Multi-Domain Systems on Chips 677 Ribbon movement vs. 1st mode power efficiency λ/4 1.2 Power efficiency 1.0 (au) 0.8 0.6 0.4 0.2 0.0 0 50 100 150 200 Ribbon movement (nm) FIGURE 20.23 Transient analysis of ribbon movement and intensity contours. The direct integration technique takes approximately 5.5 days to finish. If N is reduced to 1024, the simulation completes in approximately 25 s, whereas the direct integration simulation takes approximately 32 h. These simulations were run on a 1.7 GHz dual-processor PC running Linux, with 2 GB of main memory. In the next simulation, we perform a transient sweep of the ribbon move- ment, from 0 to 150 nm. The rest of the system setup is exactly the same as before. However, this time, we simulate the normalized power efficiency captured in the 1st diffraction mode for different ribbon depths. To simu- late this, a circular detector (radius = 12.5 μm) is placed on the positive 1st mode. Figure 20.23 is a graph that shows the simulated normalized power efficiency in this first mode. As the ribbons are moved downward, more opti- cal power is diffracted into the nonzero modes. As the ribbons reach the λ/4 point, almost all the diffractive power is in the +1st mode. Figure 20.23 also includes intensity contours of selected wave fronts during the transient sim- ulation, along with the markings of the system origin and circular detector position. From these wave fronts, interesting diffractive effects can be noted. As expected, when there is little or no ribbon movement, all the light is in the 0th mode. However, with a little ribbon movement, it is interesting to note that the 0th mode is “steered” at a slight angle from the origin. As the ribbons move downward about λ/8, the energy in the +1st modes are clearly defined. As the gratings move closer to the λ/4 point, the power is shifted from the 0th mode into the +1st modes, until there is a complete switch. As the ribbons move past the λ/4 point, optical power shifts back into the 0th mode. In the final simulation, we present a full system-level example as we expand the system to show a complete end-to-end link used in a config- uration of a color projection system. The system is shown in Figure 20.24.
678 Model-Based Design for Embedded Systems 1000 μm Prism Screen (70 μm) Detector GLV Color wheel Lens ( f = 500 μm) Input light FIGURE 20.24 End-to-end GLV display link. In this system, we model light, passing through a color wheel, striking a prism, reflecting off the GLV device, past a screen, focused by a lens, and striking a detector [44]. In this system, when the GLV ribbons are all up, the screen blocks the light’s 0th mode and the pixel is not displayed. When the alternating ribbons are pulled down, the lens focuses the light found in the +1st modes and converges them to the center of the system, display- ing the pixel. Using a spinning color wheel to change the wavelength of the incident light, a frame-sequential GLV projection system uses red (680 nm), green (530 nm), and blue (470 nm) light on the same grating. Since the same grating is used for all wavelengths of light, the grating movement is tuned for the middle frequency: 130 nm (λgreen /4). During this simulation, we use a hybrid approach for the optical modeling. For the propagation through the color wheel and the prism, we use Gaussian propagation. Since propa- gating through these components does not diffract the beam, this Gaussian technique is not only efficient, but valid. However, as soon as the light prop- agates past the prism component, we switch the optical propagation tech- nique to our full scalar method to accurately model the diffraction off the GLV device. The remainder of the simulation is propagated with the scalar technique. We analyze the system by looking at the amount of optical power that is being received on a centered circular detector (radius 10 μm) for the different wavelengths of light, since we are using the same GLV that is tuned for the green wavelength for all wavelengths. A sweep of the distance between the focusing lens and the detector plane is simulated for 0–1500 μm, when the GLV ribbons are pulled down. The graph in Figure 20.25 shows the normalized power received on the circular detector for each wavelength along with selected intensity contours of the green wave front as the beam propagates past the lens. For clarity, the detector’s size and position is added onto the intensity contours. For distances under 600 μm, the light remains in
CAD Tools for Multi-Domain Systems on Chips 679 Normalized power efficiency vs. distance between lens and detector plane 1.2 –5e – 05 1.0 0 5e – 05 Optical efficiency (au) 0.8 –5e – 05 0.6 0 5e – 05 0.4 –5e – 05 0 0.2 5e – 05 0.0 0 500 1000 1500 Distance between lens and detector (μm) Green Red Blue FIGURE 20.25 Wavelength power versus distance propagated. its two positive and negative 1st modes, as the convergence of the beams has not occurred, resulting in zero power being received on the center detector. As expected, each of the wavelengths focuses at a different rate, as shown by each wavelength’s specific curve in Figure 20.25. However, it is seen that all wavelengths focus and achieve detected maximum power at a distance past the lens of 1000 μm, or twice the lens’ focal length. At this point, all three colors project on top of each other, creating a color pixel in the focal plane. With additional optics, this focal plane can be projected to a screen outside the projector. This simulation has shown that the grating, although tuned for the green wavelength, can be used for all three wavelengths. Having shown the use of Chatoyant for modeling multi-domain ana- log systems, we now turn to the problem of co-simulation between the framework described above and a traditional HDL simulator. Co-simulation requires the solution of two problems at the interface between the simula- tors. First, a consistent model of time must be reached for when events occur. Second, a consistent model of signal values must be developed for signals crossing the interface. This is the subject of the next section. 20.3 HDL Co-Simulation Environment The two levels of simulation discussed above, component and analog system that are supported by Chatoyant, have not been optimized to
680 Model-Based Design for Embedded Systems simulate designs that are specified in an HDL such as Verilog or VHDL. There are no components in the Chatoyant library that directly use HDL as an input language. On the other hand, there are many available commer- cial and research mixed-language HDL simulators. Mixed-language refers to the ability for a simulator to compile and execute VHDL, Verilog, and Sys- temC (or other C/C++ variants). In an earlier work we investigated the use of CoSim with Chatoyant models [47]. In this section, we explore an interface to a commercial system. Cadence, Mentor Graphics, Synopsys, and other EDA companies provide such simulators. One common feature among the more widely used simulators, such as ModelSim and NCSIM, is the abil- ity to execute C-based shared object files embedded in HDL design objects. These simulators provide an application programmer’s interface (API) to gain access to simulator data and control design components. ModelSim was chosen since it has a large set of C routines that allow access to sim- ulator state as well as modifying design signals and runtime states. These functions and procedures are bundled in an extension package known as the foreign language interface (FLI) [48]. By creating a co-simulation envi- ronment between ModelSim and Chatoyant, a powerful MDSoC design and verification environment has been created. This environment is able to address the demand for a robust and efficient system architecture/ design space exploration and prototyping tool that can support the design of MDSoCs. The rest of this chapter focuses on the development of the interface between Chatoyant and ModelSim and the performance of the resulting environment. 20.3.1 Architecture The architecture of the co-simulation environment is kept simple to be as efficient and accurate as possible. There are two phases to the execution of the environment: a system generation phase and a runtime support environ- ment. Each is a standalone process, but both are required for system simula- tion. Figure 20.26 illustrates this top-level structure. 20.3.1.1 System Generator The System Generator allows the user to create the necessary files needed by both Chatoyant and ModelSim. For Chatoyant this includes a common header and object file used in both simulators as well as components (stars) used for the Chatoyant side of the interface. The same header and object file are used for ModelSim, in addition to a shared object library file that is used for invoking the ModelSim FLI when ModelSim is loaded and elaborates a design. The main input to this generator is the top-level or interface-specific VHDL file. This file contains the list of ports that represent the main conduit
CAD Tools for Multi-Domain Systems on Chips 681 Top - level VHDL file System generator Chatoyant Definitions Wrapper FLI share star library VHDL object file Chatoyant Co-simulation runtime ModelSim system FIGURE 20.26 Co-simulation top-level structure. between the digital domain running within ModelSim and the other domains handled in Chatoyant. When this file is loaded by the System Generator, the entity portion of the VHDL is parsed and a linked list of the ports is created. Each node in this linked list contains the port’s name, its direction (in/out/bidirectional), and its width (1 bit for a signal and n bits for a bus). Using a graphical user interface, the user can select which ports to include and the mapping for the analog voltage levels to be converted into and out of the MVL9 (Multi-Value Logic 9 signal representation standard) logic repre- sentation used by ModelSim. There are four fields for this including a high, a low, a cutoff for high, and a cutoff for low voltage values. The user also specifies a name for the system, used for code generation and library man- agement. The outputs of the generator phase are the component star file for Chatoyant, the FLI source code for the ModelSim FLI, the header and source files for a common resource library for the system, a makefile for remaking the object files, a usage text file, and the first time compilation object files performed at the end of the generation. With these files in place, the user can then proceed with the execution of the linked simulators. 20.3.1.2 Runtime Environment: Application of Parallel Discrete Event Simulation The runtime system differentiates itself from other typical co-simulation environments in that there is no central simulation management system. Chatoyant and ModelSim are treated as two standalone processes and
682 Model-Based Design for Embedded Systems communicate only between themselves. This reduces the overhead of another application executing along with the two simulators as well as the additional message traffic produced by such an arbiter. This philosophy is an application of a general parallel discrete event sim- ulation (PDES) system. Since there are two standalone processes, each is treated as if it were its own DE processing node. Without a central arbiter, the two must (1) exchange event information by converting logic values into voltages and vice versa, and (2) synchronize their respective local simula- tion times. To exchange the event information, the system uses technology- specific lookup tables, created by the System Generator, that provide the conversion between a logic “1” and a logic “0” to a voltage in addition to determining what voltage level constitutes a logic “1” and “0.” The synchronization of the simulators is where the application of PDES methods enters [49]. The asynchronous DE simulation invokes both simula- tors to perform unique tasks on separate parts of a design in a nonsequential fashion. This is because of the fact that there is no master synchronization process as in [1]. For synchronization and scheduling there are two major approaches one can take, conservative or optimistic. We discuss our choice next. 20.3.1.3 Conservative versus Optimistic Synchronization The conservative and optimistic approaches solve the parallel synchroniza- tion problem in two distinct ways. This problem is defined in [2] as the requirement for multiple processing elements to produce events of an equal timestamp in order to not violate the physical causality of the system. The conservative method solves this problem by constraining each processing node to remain in synchronicity with the others, never allowing one simula- tor’s time to pass any other simulator. This can have the penalty of reducing the performance of a simulation by requiring extra overhead in the form of communication and deadlock avoidance. The optimistic approach breaks the rule of maintaining strict causality by allowing each processing element to simulate without considering time in other processing element. This means that the simulators can run freely without having to synchronize, with the exception of communicating explicit event information. If, however, there is an event sent from one simulator to the other, and the second simulator has a local current time greater than the event’s timestamp, then the receiving simulation process must stop and roll- back time to a known safe state that is before the timestamp of the incoming event. This approach requires state saving as well as rollback mechanisms. This can be costly in terms of memory usage and processing overhead for determining and recalling previous states, and thus increases the processing time of every event. Both approaches are possible since ModelSim does have check-pointing and restoring methods available [48]. However, the conservative PDES
CAD Tools for Multi-Domain Systems on Chips 683 method was chosen as the underlying philosophy for our co-simulation solution. Two factors went into this decision. The first consideration is that the co-simulation environment is executing as two processes on one workstation, so that exchanging timing information is not as costly as in a large physically distributed simulation environment. The second is that even with a dual-processor workstation, there is not an excess of computational or memory resources that is seen in a truly distributed PDES architecture, and therefore, a rollback would be too costly. This was confirmed with a preliminary test of the fiber image guide sys- tem described below. For that system the amount of data required for a checkpoint file was on the order of 1 to 2MB. With an average of 10 check- point files needed to keep the two simulators within a common time horizon, rollback time took between 500 ms and 1.5 s. On the other hand, the conservative approach gives a solution requir- ing significantly less memory at the expense of increased communication to ensure that both simulators are consistently synchronized. This becomes a matter of passing simple event time information between the two simulators. Thus, the only real design issue becomes the time synchronization method. 20.3.1.4 Conservative Synchronization Using UNIX IPC Mechanisms As described in more detail below, the system was developed and tested on a Linux-based workstation. Therefore, UNIX-style IPC is used for the commu- nication architecture. Event information is exchanged using shared memory, and synchronization is achieved by using named pipes in blocking mode. This is similar to the synchronized data transfer and blocking methodology described in [50]. With these two mechanisms, the conservative approach is implemented in the two algorithms seen in Figure 20.27. The algorithm for the co-simulation is straightforward. Both simulators, running concurrently, reach a point in their respective execution paths where they enter the interface code in Figure 20.27. Both check to ensure that they are at the next synchronization point (next_sync), and if they are not, they exit this section of code and continue. If they are at the next synchroniza- tion point, defining the safe-point in terms of the conservative approach in PDES, then Chatoyant starts the exchange by checking for any change in its outputs to ModelSim. If there is any change in any bit of these ports, that port is marked dirty, and a change flag is set. When all the ports have been examined, Chatoyant sends ModelSim either a ModelSim_Bound event, if any port changed value, or a No_Change event. Simultaneously, ModelSim waits for this event message from Chatoyant. Once received, it will update and schedule an event for those ports with dirty flags set, if any. It then jumps to check its own output ports, checking bit by bit for a change in each port’s value. Once again, as in Chatoyant, if there is a difference, the dirty flag for that port is set, and the change flag in ModelSim is set true. Once this is done for every port, ModelSim will send a message to Chatoyant that there is either a change (Chatoyant_Bound) or No_Change.
684 Model-Based Design for Embedded Systems Chatoyant ModelSim If(time < next_sync) If(time < next_sync) return at a later time; return at a later time; For each output: Wait(Chatoyant_Response); For each bit in signal: If(cur[i] ! = new[i]) If(Response = = No_Change) mark dirty; ; flag change; Else End If; For each input: End For each bit; If(input.dirty) End For each output; update local value; ScheduleEvent(); If(change){ clear input.dirty; send(ModelSim_Bound); End If; Else End For each input; send(No_Change); End If; end If; For each output: Wait(ModelSim_Response); // Blocking For each bit in signal: If(cur[i] ! = new[i]) If(Response == No_Change) mark dirty; goto Synchronize; flag change; Else End If; For each input: End For each bit; If(input.dirty) End For each output; update local value; ScheduleEventToPorthole(); If(change){ clear input.dirty; send(Chatoyant_Bound); End If; Else End for each input; send(No_Change); End If; end If; Synchronize: Synchronize: next_sync = now + SYNC_PULSE; next_sync = now + SYNC_PULSE; Send(Chatoyant_Finished); Wait(Chatoyant Finished); Wait(ModelSim_Finished); Send(ModelSim_Finished); Done with iteration; Done with iteration; FIGURE 20.27 The synchronization in both simulators. Chatoyant, waiting for this response, will receive it and take action sim- ilar to that of ModelSim in updating the inputs from ModelSim. Finally, the two will set their respective next synchronization times and handshake with one another to indicate it is safe to continue simulating. The No_Change mes- sages are analogous to the null message passing scheme defined by Chandy and Misra [49], which has the benefit of avoiding simulation deadlock. A key point is the concept of the next synchronization time (next_sync). This value is calculated based on a global parameter in the co-simulation
CAD Tools for Multi-Domain Systems on Chips 685 environment known as the SYNC_PULSE. This parameter defines the resolution of how often synchronization occurs. This value ultimately defines the speed versus accuracy tradeoff ratio between the simulators. A higher resolution (smaller SYNC_PULSE value) means greater accuracy but slower runtime. Depending on a particular system, this could affect the quality of the simulation results. 20.3.2 Co-Simulation of Experimental Systems To examine the effects of synchronization resolution on speed and accuracy, we simulate two example MDSoC systems. Both are large-scale systems, meaning there are many components in each domain, including multi- ple analog circuits, complex optics, and mixed wire and bus interconnects between the digital and analog domains. 20.3.2.1 Fiber Image Guide The first of these systems is the fiber image guide, or FIG, system developed at the University of Pittsburgh [51]. FIG is a high-speed 64 × 64-bit opto- electronic crossbar switch built using an optical multi-chip module. FIG uses guided wave optics, analog amplification and filtering circuits, and digital control logic to create an 8 × 8, 8-bit bus crossbar switch. The switch is built as a multistage interconnection network (MIN) built with a shuffle-exchange architecture. The shuffle operations are performed by the wave guide, and the digital logic performs the exchange switching operation. Analog circuits amplify the digital signals and drive VCSEL arrays which in turn transmit light through the image guide. Photodetectors are used to convert the light back into an analog signal, which is amplified and fed back into the digital domain. This system, illustrated in Figure 20.28, exercises the ability of the co-simulation environment to handle buses as well as the communications between domains without a synchronous clock. In other words, there is no clock signal traveling across the co-simulation interface, and thus the events occur in asynchronous fashion. 20.3.2.2 Smart Optical Pixel Transceiver The smart optical pixel transceiver, or SPOT, was a development at the Uni- versity of Delaware [52]. It provides a short-range free-space optical link between two custom-designed transceivers. Each transceiver either accepts or generates a parallel bus, in the digital domain. On the transmitter side, each bus is serialized into a double data rate data signal, along with a 4X clock (125 MHz clock doubled to 250 MHz in this test system). Serializa- tion and de-serialization are handled in the digital domain. These serial data/clock streams are converted into analog signals that are amplified and used to drive VCSEL arrays, similar to FIG. Photodetectors convert the