OpenLANE-Sky130-Physical-Design-Workshop
This is the compilation of my notes for the 5 Day Workshop: Advanced Physical Design using OpenLANE/Sky130 from August 3-7, 2022. The goal is to cover the complete RTL2GDS flow using the open-source flow OpenLane with SKY130nm PDK.
Table of Contents
DAY 1: Inception of Open-source EDA, OpenLane and Sky130 PDK
The core of the chip will contain two types of blocks:
- Foundry IP Blocks (e.g. ADC, DAC, PLL, and SRAM) = blocks which requires some amount of intelligent techniques to build which can only be designed by foundries.
- Macro blocks (e.g. RISC-V SOC and SPI) = pure digital logic blocks compared to IP’s which might require some analog parts.
Open Source Digital ASIC Design requires three open-source components:
PDK (Process Design Kit) = A set of data files and documents which serves as the interface between the designer and the fab. This includes cell libraries, IO libraries, design rules (DRC, LVS, etc.)
Simplified RTL to GDSII Flow:
- Sythesis = The RTL is converted into a gate level netlist made up of components of standard cell libary.
- Floor Planning/ Power Planning = Plan silicon area and create robust power distribution network. The power network usually uses the upper metal layer which are thicker than lower layer and thus lower resistance. This lowers the IR drop problem
- Placement = There are two steps, first is global placement which is the general optimal positons for cells and might not be legal. Next is detailed placement which is the actual legal placements of the cells.
- Clock tree synthesis = clock distribution to all flip flops and is usually a tree (H-tree, X-tree … )
- Routing = Use horizontal and vertical wires to connect cells together. The router uses PDK information (thickness, pitch, width,vias) for each metal layer to do the routing. The Sky130 defines 6 routing layers. It doe global routing and detailed routing.
- Verification before sign-off = Involves physical verification like DRC and LVS and timing verification. Design Rule Checking or DRC ensures final layout honors all design rules and Layout versus Schematic or LVS ensures final layout matches the gate level netlist from synthesis phase. Timing verification ensures timing constraints are met.
The final layout is in GDSII file format.
OpenLane = An open-source ASIC development flow reference. It consists of multiple open-source tools needed for the whole RTL to GDSII flow. This is tuned epecially for Sky130 PDK. It also works for OSU 130nm. It is recommended to read the OpenLANE documentation before moving forward.
-
The input for the whole flow are the rtl files, sdc file, and PDK files. The output is GDSII/LEF file.
-
Yosys is used to convert the HDL to gate level netlist using generic components. The ABC script is then used to map the generic components to the standard cell library of the PDK. These ABC scripts is used to make various synthesis strategies (using the Synthesis Exploration) which can optimize the design either with least area or best timing.
-
The Logic Equivalency Cheking (LEC) is used to compare the resulting netlist after optimization of place and route to the gate level netlist from synthesis phase
-
Antenna Rules Violation = long wire segments will act as antennna and will accumulate charges, this might damage the connected transistor gates. Solution is to either use bridging or antenna diode insertion to leak away the charges
OpenLane Directory Hierarchy:
├── OOpenLane -> directory where the tool can be invoked (run docker first)
│ ├── designs -> All designs must be extracted from this folder
│ │ │ ├── picorv32a -> Design used as case study for this workshop
│ | | ├── ...
| | ├── ...
├── pdks -> contains pdk related files
│ ├── skywater-pdk -> all Skywater 130nm PDKs
│ ├── open-pdks -> contains scripts that makes the commerical PDK (which is normally just compatible to commercial tools) to also be compatible with the open-source EDA tool
│ ├── sky130A -> pdk variant made especially compatible for open-source tools
│ │ │ ├── libs.ref -> files specific to node process (timing lib, cell lef, tech lef) for example is `sky130_fd_sc_hd` (Sky130nm Foundry Standard Cell High Density)
│ │ │ ├── libs.tech -> files specific for the tool (klayout,netgen,magic...)
Inside a specific design folder contains a config.tcl which overrides the default settings on OpenLANE. These configurations are specific to a design (e.g. clock period, clock port, verilog files…). The priority order for the OpenLANE settings:
- sky130_xxxxx_config.tcl in
OpenLane/designs/[design]/
- config.tcl in
OpenLane/designs/[design]/
- Default values in
OpenLane/configuration/
Lab [Day 1] - Determine Flip-flop Ratio:
The task is to find the flip-flop ratio ratio for the design picorv32a. This is the ratio of the number of flip flops to the total number of cells. For the OpenLane installation, the steps are very straight forward and can be found on the OpenLane repo.
1. Run OpenLANE:
$ make mount = Open the docker platform inside the openlane/% flow.tcl -interactive = run script for automating the whole RTL to GDSII flow but in step by step -interactive mode% package require openlane 0.9 == retrives all dependencies for running v0.9 of OpenLANE
2. Design Setup Stage:
% prep -design picorv32a = Setup the filesystem where the OpenLANE tools can dump the outputs. This also creates a run/ folder inside the specific design directory which contains the command log files, results, and the reports dump by each tools. These folders will be empty for now except for lef files generated by this design setup stage. This merged the cell LEF files .lef and technology LEF files .tlef generating merged.nom.lef inside run/tmp/
3. Run synthesis:
% run_synthesis = Run yosys RTL synthesis, ABC scripts (for technology mapping), and OpenSTA.
The flipflop ratio is (number of flip flops)/(total number of cells) is 1613/14876 = 0.10843. Or 10.843%
After running synthesis, inside the runs/[date]/results/synthesis is picorv32a_synthesis.v which is the mapping of the netlist to standard cell library using ABC. The runs/[date]/reports/synthesis will contain synthesis statistic reports and static timing analysis reports. The runs/[date]/synthesis/logs contains log files for the terminal output dumps for running yosys and OpenSTA.
DAY 2: Good Floorplan vs Bad Floorplan and Introduction to Library Cells
Floorplan Stage:
-
Find height and width of core and die.
Core is where the logic blocks are placed and this seats at the center of the die. The width and height depends on dimensions of each standard cells on the netlist. Utilization factor is (area occupied by netlist)/(total area of the core). In practical scenario, utilization factor is 0.5 to 0.6. This is space occupied by netlist only, the remaining space is for routing and more additional cells. Aspect ratio is (height)/(width) of core, so only aspect ratio of 1 will produce a square core shape.
-
Define location of Preplaced Cell.
These are reusable complex logicblocks or modules or IPs or macros that is already implemented (memory, clock-gating cell, mux, comparator…) . The placement on the core is user-defined and must be done before placement and routing (thus preplaced cells). The automated place and route tools will not be able to touch and move these preplaced cells so this must be very well defined
-
Surround preplaced cells with decoupling capacitors.
The complex preplaced logicblock requires a high amount of current from the powersource for current switching. But since there is a distance between the main powersource and the logicblock, there will be voltage drop due to the resistance and inductance of the wire. This might cause the voltage at the logicblock to be not within the noise margin range anymore (logic is unstable). The solution is to use decoupling capacitors near the logic block, this capacitor will send enough current needed by the logicblock to switch within the noise margin range.
-
Power Planning
Decoupling capactor for sourcing logic blocks with enough current is not feasible to be applied all over the chip but only on the critical elements (preplaced complex logicblocks). Large number of elements switching to logic 0 might cause ground bounce due to large amount of current that needs to be sink at the same time, and switcing to logic 1 might cause voltage droop due to not enough current from the powersource to source needed current of all elements. Ground bounce and voltage droop might cause the voltage to not be within the noise margin range. The solution is to have multiple powersource taps (power mesh) where elements can source current from the nearest VDD and sink current to the nearest VSS tap. This is the reason why most chips have multiple powersource pins.
-
Pin Placement
The input and output ports are placed on the space between the core and the die. The placements of the ports depens on where the cells connected to those ports are placed on the core. The clock port is thicker(least resistance path) than data ports since this clock must be capable to drive the whole chip.
-
Logical Cell Placement Blockage
This makes sure that the automated placement and routing tool does not place any cell on the pin locations of the die.
Below are all 6 steps for floor planning:
Placement Stage:
- Bind the netlist to a physical cell with real dimensions. The physical cell will come from a library that can provide multiple options for shapes, dimensions, and delay for same cells.
- Next is placement of those physical cells to the floorplan. The flip flops must be placed as near as possible to the input and output pins to reduce timing delay.
- Optimize placement to maintain signal integrity. This is where we estimate wirelength and capacitance (C=EA/d) and based on that insert repeaters/buffers. The wirelength will form a resistanace which will cause unnecessary voltage drop and a capacitance which will cause a slew rate that might not be permissible for fast current switching of logic gates. The solution to reduce resistance and capacitance is to insert buffers for long routes that will act as intermediary and separate a single long wire to multilple ones. Sometime we also do abutment where logic cells are placed very close to each other (almost zero delay) if it has to run at high frequency (2GHz). Crisscrossing of routes is a normal condition for PnR since we can use separate metal layer (using vias) for crisscrossed path.
- After placement optimization, We will setup timing analysis using idle clock (zero delay for wires and has no clock buffer related delays) considering we have not yet done CTS.
The goal of placement is not yet on timing but on congestion. Also, standard cells are not placed on floorplan stage, it is done on Placement stage. Macros or preplaced cells are the ones placed on floorplan stage.Macros or preplaced cells are placed on floorplan stage.
Placement is done on two stages:
- Global Placement = placement with no legalizations and goal is to reduce wirelength. It uses Half Perimeter Wirelength (HPWL) reduction model.
- Detailed Placement = placement with legalization where the standard cells are placed on stadard rows, abutted, and must have no overlaps
Lab [Day 2] - Determine Die Area:
1. Set configuration variables. Before running floorplan stage, the configuration variables or switches must be configured first. The configuration variables are on openlane/configuration:
.
├── README.md
├── checkers.tcl
├── cts.tcl
├── floorplan.tcl
├── general.tcl
├── lvs.tcl
├── placement.tcl
├── routing.tcl
└── synthesis.tcl
The README.md describes all configuration variables for every stage and the tcl files contain the default OpenLANE settings. All configurations accepted by the current run is on openlane/designs/picorv32a/runs/config.tcl. This may come either from (with priority order):
- PDK specific configuration inside the design folder
config.tcl inside the design folder- System default settings inside
openlane/configurations
2. Run floorplan on OpenLane: % run floor_plan
3. Check the results. The output of this stage is runs/[date]/results/floorplan/picorv32a.floorplan.def which is a design exchange format, containing the die area and positions.
...........
DESIGN picorv32a ;
UNITS DISTANCE MICRONS 1000 ;
DIEAREA ( 0 0 ) ( 660685 671405 ) ;
............
The die area here is in database units and 1 micron is equivalent to 1000 database units. Thus area of the die is (660685/1000)microns*(671405/1000)microns = 443587 microns squared.
4. View the layout on magic. Open def file using magic:
magic -T /home/kunalg123/Desktop/work/tools/openlane_working_dir/pdks/sky130A/libs.tech/magic/sky130A.tech lef read ../../tmp/merged.lef def read picorv32a.floorplan.def
To center the view, press “s” to select whole die then press “v” to center the view. Point the cursor to a cell then press “s” to select it, zoom into it by pressing 'z". Type “what” in tkcon to display information of selected object. These objects might be IO pin, decap cell, or well taps as shown below.
Useful Magic commands are listed on the Magic Commands section.
5 Run placement: % run_placement. This commmand is a wrapper which does global placement (performed by RePlace tool), Optimization (by Resier tool), and detailed placement (by OpenDP tool). It displays hundreds of iterations displaying HPWL and OVFL. The algorithm is said to be converging if the overflow is decreasing. It also checks the legality.
6. View the output of this stage. The output of this stage is runs/[date]/results/placement/picorv32a.placement.def. To see actual layout after placement, open def file using magic:
magic -T /home/kunalg123/Desktop/work/tools/openlane_working_dir/pdks/sky130A/libs.tech/magic/sky130A.tech lef read ../../tmp/merged.lef def read picorv32a.placement.def
Library Characterization:
Of all RTL-to-GDSII stages, one common thing that the EDA tool always need is data from the library of gates which keeps all standards cells (and, or, buffer gates,…), macros, IPs, decaps, etc. Same cells might have different flavors inside the library (different sizes, delays, threshold voltage). Bigger cell sizes means bigger drive strength to drive longer and thicker wires. Bigger threshold voltage (due to bigger size) will take more time to switch(slower clock) than those with smaller threshold voltage.
A single cell needs to go through the cell design flow. The inputs to make a single cell comes from the foundry Process Design Kits:
- DRC & LVS Rules = tech files and poly subtrate paramters (CUSTOME LAYOUT COURSE)
- SPICE Models = Threshold, linear regions, saturation region equations with added foundry parameters. Including NMOS and PMOS parameteres (Ciruit Deisgn and Spice simulation Course)
- User defined Spec = Cell height (separation between power and ground rail), Cell width (depends on drive strength), supply voltage, metal layer requirement (which metal layer the cell needs to work)
The library cell developer must adhere to the rules given on the inputs so that when the cell is used on a real design, there will be no errors. Next is design the library cell:
- Design the circuit function (Output: circuit design language (CDL))
- Model the pmos and nmos that meets input library requirement
- Layout the design using Euler’s path and sticky diagram to produce best area. This can be done on
magic layout tool.The outputs are:
- GDSII (layout file)
- LEF (defines the width and height of cell)
- extract spice netlist .cir (parasitics of each element of cell: resistance, capacitance)
Afte design is characterization using GUNA software, where the outputs are timing, noise, and power characterization.
.
Timing Characterization:
Below are the timing variables for slew. This is two inverters in series, red is output of first inverter and blue is output of second inverter:
Below are the timing variables for propagation delay. The red is input waveform and blue is output waveform of the buffer. The left side is rise delay and right side is fall delay.
Negative propagation delay is unexpected. That means the output comes before the input so designer needs to choose correct threshold point to produce positive delay. Delay threshold is usually 50% and slew rate threshold is usually 20%-80%.
DAY 3: Design a Library Cell using Magic Layout and Ngspice Characterization
Configurations on OpenLANE can be changed on the flight. For example, to change IO_mode to be not equidistant, use % set ::env(FP_IO_MODE) 2; on OpenLANE. The IO pins will not be equidistant on mode 2 (default of 1). Run floorplan again via % run_floorplan and view the def layout on magic. However, changing the configuration on the fly will not change the runs/config.tcl, the configuration will only be available on the current session. To echo current value of variable: echo $::env(FP_IO_MODE)
Designing a Library Cell:
- SPICE deck = component connectivity (basically a netlist) of the CMOS inverter.
- SPICE deck values = value for W/L (0.375u/0.25u means width is 375nm and lengthis 250nm). PMOS should be wider in width(2x or 3x) than NMOS. The gate and supply voltages are normally a multiple of length (in the example, gate voltage can be 2.5V)
- Add nodes to surround each component and name it. This will be used in SPICE to identify a component.
Notes:
- Width is the length of source and drain. Length is the distance between source and drain
- PMOS’ hole carrier is slower than NMOS’ electron carrier mobility, so to match the rise and fall time PMOS must be thicker (less resistance thus higher mobility) than NMOS
- A good refresher on MOSFETS and CMOS is this video and this site.
SPICE Deck Netlist Description:
Notes:
- Syntax for the PMOS and NMOS descriptiom:
[component name] [drain] [gate] [source] [substrate] [transistor type] W=[width] L=[length]
- All components are described based on nodes and its values
.op is the start of SPICE simulation operation where Vin will be sweep from 0 to 2.5 with 0.5 stepstsmc_025um_model.mod is the model file containing the technological parameters for the 0.25um NMOS and PMOS
The steps to simulate in SPICE:
source [filename].cir
run
setplot
dc1
plot out vs in
SPICE Analysis for Switching Threshold and Propagation Delay:
CMOS robustness depends on:
-
Switching threshold = Vin is equal to Vout. This the point where both PMOS and NMOS is in saturation or kind of turned on, and leakage current is high. If PMOS is thicker than NMOS, the CMOS will have higher switching threshold (1.2V vs 1V) while threshold will be lower when NMOS becomes thicker.
-
Propagation delay = rise or fall delay
DC transfer analysis is used for finding switching threshold. SPICE DC analysis below uses DC input of 2.5V. Simulation operation is DC sweep from 0V to 2.5V by 0.05V steps:
Vin in 0 2.5
*** Simulation Command ***
.op
.dc Vin 0 2.5 0.05
Below is the result of SPICE simulation for DC analysis, the line intersection is the switching threshold:
Meanwhile, transient analysis is used for finding propagation delay. SPICE transient analysis uses pulse input:
- starts at 0V
- ends at 2.5V
- starts at time 0
- rise time of 10ps
- fall time of 10ps
- pulse-width of 1ns
- period of 2ns
The simulation operation has 10ps step and ends at 4ns:
Vin in 0 0 pulse 0 2.5 0 10p 10p 1n 2n
*** Simulation Command ***
.op
.tran 10p 4n
Below is the result of SPICE simulation for transient analysis:
CMOS Fabrication Process (16-Mask CMOS Process):
1. Selecting a substrate = Layer where the IC is fabricated. Most commonly used is P-type substrate
2. Creating active region for transistor = Separate the transistor regions using SiO2 as isolation
- Mask 1 = Covers the photoresist layer that must not be etched away (protects the two transistor active regions)
- Photoresist layer = Can be etched away via UV light
- Si3N4 layer = Protection layer to prevent SiO2 layer to grow during oxidation (oxidation furnace)
- SiO2 layer = Grows during oxidation (LOCOS = Local Oxidation of Silicon) and will act as isolation regions between transistors or active regions
3. N-Well and P-Well Fabrication = Fabricate the substrate needed by PMOS (N-Well) and NMOS (P-Well)
- Phosporus (5 valence electron) is used to form N-well
- Boron (3 valence electron) is used to form P-Well.
- Mask 2 protects the N-Well (PMOS side) while P-Well (NMOS side) is being fabricated then Mask 3 while N-Well (PMOS side) is being fabricated
4. Formation of Gate = Gate fabrication affects threshold voltage. Factors affecting threshold voltage includes:
Main parameters are:
- Doping Concentration = Controlled by ion implantation (Mask 4 for Boron implantation in NMOS P-Well and Mask 5 for Arsenic implantation in PMOS N-Well)
- Oxide capacitance = Controlled by oxide thickness (SiO2 layer is removed then rebuilt to the desire thickness)
Mask 6 is for gate formation using polysilicon layer.
5. Lightly Doped Drain formation = Before forming the source and drain layer, lightly doped impurity is added:
- Mask 7 for N- implantation (lightly doped N-type) for NMOS
- Mask 8 for P- implantation (lightly doped P-type) for PMOS.
Heavily doped impurity (N+ for NMOS and P+ for PMOS) is for the actual source and drain but the lightly doped impurity will help maintain spacing between the source and drain and prevent hot electron effect and short channel effect.
6. Source and Drain Formation = Mask 9 is for N+ implantation and Mask 10 for P+ implantation
- Channeling is when implantations dig too deep into substrate so add screen oxide before implantation
- The side-wall spacers maintains the N-/P- while implanting the N+/P+
7. Form Contacts and Interconnects = TiN is for local interconnections and also for bringing contacts to the top. TiS2 is for the contact to the actual Drain-Gate-Source. Mask 11 is for etching off the TiN interconnect for the first layer contact.
8. Higher Level Metal Formation = We need to planarize first the layer via CMP before adding a metal interconnect. Aluminum contact is used to connect the lower contact to higher metal layer. Process is repeated until the contact reached the outermost layer.
- Mask 12 is for first contact hole
- Mask 13 is for first Aluminum contact layer
- Mask 14 is for second contact hole
- Mask 15 is for second Aluminum contact layer. Mask 16 is for making contact to topmost layer.
When polysilicon crosses N-diffusion/P-diffusion (diffusion is also called implantation), then an NMOS/PMOS is created. Explained here is the reason why the diffusion layer of source and drain “seems” to be connected under the polysilicon (diffusion layer for source and drain supposedly be separated).
The first layer is local-interconnect layer or local-i then metal 1 to 5. Here is the process stack diagram of sky130nm PDK. Metal 1 is for Power and Ground lines. Nsubstratecontact connects the N-well to locali. licon connects the locali to metal1.Locali is for local connections of cells.
The layer hierarchy for NMOS is: Psubstrate -> Psubstrate Diffusion (psd) -> Psubstrate Contact (psc) -> Local-interconnect (li) -> Mcon -> Metal1. For poly: Poly -> Polycontact -> Locali. P-substrate diffusion an N-substrate diffusion is also referred to as P-tap and N-tap.
The output of the layout is the LEF file. LEF (Library Exchange Format) is used by the router tool in PnR design to get the location of standard cells pins to route them properly. So it is basically the abstract form of layout of a standard cell. picorv32a/runs/[DATE]/tmp contains the merged lef files (cell LEF and tech LEF). Notice how metal layer directon (horizontal or vertical) is alternating. Also, metal layer width and thickness is increasing.
Magic Commands:
Here is a great video guide on layout using Magic. And here is the Magic website with tutorials.
- Left click = lower-left corner of box
- Right click = upper-right corner of box
- “z” = zoom in, “Z” = zoom out, “ctrl + z” = zoom into the box
- Middle click on empty area will turn the box into empty (similar to erasing it)
- “s” three times will select all geometries electrically connected to each other
:box = display parameters of selected box:grid 0.5um 0.5um = turn on/off and set grid:snap user = snap based on current grid:help snap = display help for command:drc style drc(full) = use all DRC when doing DRC checking:paint poly = paint “poly” to current box:drc why = show drc violation inside selected area (white dots are DRC violations ):erase poly = delete poly inside the box:select area = select all geometries inside the box:copy n 30 = copy selected geometries to North by 30 grid steps:move n 1 = move selected geometries to North by 1 step (“.” to move more, “u” to undo): select cell _08555_ = select a particular cell instance (e.g. cell _08555_ which can be searched in the DEF file):cellname allcells = list all cells in the layout:cellname exists sky130_fd_sc_hd__xor3_4 = check if a cell exists:drc why = show DRC violation and also the DRC name which can be referenced from Sky130 PDK Periphery Rules.
Lab Part 1 [Day 3] - Slew Rate and Propagation Delay Characterization:
The task is to characterize a sample inverter cell by its slew rate and propagation delay.
-
Clone vsdstdcelldesign. Copy the techfile sky130A.tech from pdks/sky130A/libs.tech/magic/ to directory of the cloned repo. Below are the contents of vsdstdcelldesign/libs/:
-
View the mag file using magic magic -T sky130A.tech sky130_inv.mag &:
- Make an extract file
.ext by typing extract all in the tkon terminal.
- Extract the
.spice file from this ext file by typing ext2spice cthresh 0 rthresh 0 then ext2spice in the tcon terminal.
We then modify the spice file to be able to plot a transient response:
* SPICE3 file created from sky130_inv.ext - technology: sky130A
.option scale=0.01u
.include ./libs/pshort.lib
.include ./libs/nshort.lib
* .subckt sky130_inv A Y VPWR VGND
M0 Y A VGND VGND nshort_model.0 ad=1435 pd=152 as=1365 ps=148 w=35 l=23
M1 Y A VPWR VPWR pshort_model.0 ad=1443 pd=152 as=1517 ps=156 w=37 l=23
C0 A VPWR 0.08fF
C1 Y VPWR 0.08fF
C2 A Y 0.02fF
C3 Y VGND 0.18fF
C4 VPWR VGND 0.74fF
* .ends
* Power supply
VDD VPWR 0 3.3V
VSS VGND 0 0V
* Input Signal
Va A VGND PULSE(0V 3.3V 0 0.1ns 0.1ns 2ns 4ns)
* Simulation Control
.tran 1n 20n
.control
run
.endc
.end
Open the spice file by typing ngspice sky130A_inv.spice. Generate a graph using plot y vs time a :
Using this transient response, we will now characterize the cell’s slew rate and propagation delay:
-
Rise Transition [output transition time from 20%(0.66V) to 80%(2.64V)]:
- Tr_r = 2.19981ns - 2.15739ns = 0.04242 ns
-
Fall Transition [output transition time from 80%(2.64V) to 20%(0.66V)]:
- Tr_f = 4.0672ns - 4.04007ns = 0.02713ns
-
Rise Delay [delay between 50%(1.65V) of input to 50%(1.65V) of output]:
- D_r = 2.18197ns - 2.15003ns = 0.03194ns
-
Fall Delay [delay between 50%(1.65V) of input to 50%(1.65V) of output]:
- D_f = 4.05364ns - 4.05001ns =0.00363ns
Lab Part 2 [Day 3] - Fix Tech File DRC via Magic:
Read through this site about tech file. All technology-specific information comes from a technology file. This file includes such information as layer types used, electrical connectivity between types, design rules, rules for mask generation, and rules for extracting netlists for circuit simulation.
Read through also this site on the DRC rules for SKY130nm PDK
-
Download the lab contents from this site. Extract the tarball. Inside the drc_tests/ are the .mag layout files and the sky130A.tech.
-
Open magic with poly.mag as input: magic poly.mag. Focus on Incorrect poly.9 layout. As described on the poly.9 design rule of SKY130 PDK, the spacing between polyresistor with poly or diff/tap must at least be 0.480um. Using :box, we can see that the distance is 0.250um YET there is no DRC violations shown. Our goal is to fix the tech file to include that DRC.
-
Open sky130A.tech. The included rules for poly.9 are only for the spacing between the n-poly resistor with n-diffusion and the spacing between the p-poly resistor with diffusion. We will now add new rules for the spacing between the poly resistor with poly non-resistor, highlighted green below are the two added rules. On the left is the rule for spacing between n-poly resistor with poly non-resistor and on the right is the rule for the spacing between the p-poly resistor with poly non-resistor. The allpolynonres is a macro under alias section of techfile.
-
Run tech load sky130A.tech then drc check in tkcon to reload the tech file. The new DRC rules will now take effect.Notice the white dots on the poly indicating the design rule violations. Command drc find to iterate in each violations.
-
Next, notice below that there are violations between N-substrate diffusion with the polyresistors (from left: npolyres, ppolyres, xpolyres) which is good. But between npolyres with P-substrate diffusion, there is no violation shown.
-
To fix that, just modify the tech file to include not only the spacing between npolyres with N-substrate diffusion in poly.9 but between npolyres and all types of diffusion. alldif is also a macro under alias section. Load the tech file again, the new DRC will now take effect.
DAY 4: Pre-layout Timing Analysis and Importance of Good Clock Tree
To run previous flow, add tag to prep design:
prep -design picorv32a -tag [date]
PnR tool does not need all informations from the .mag file like the logic part but only PnR boundaries, power/ground ports, and input/output ports. This is what a LEF file actually contains. So the next step is to extract the LEF file from Magic. But first, we need to follow guidelines of the PnR tool for the standard cells:
- The input and output ports lies on the intersection of the horizontal and vertical tracks (ensure the routes can reach that ports).
- The width of the standard cell must be odd multiple of the tracks horizontal pitch and height must be odd multiples of tracks vertical pitch
To check these guidelines, we need to change the grid of Magic to match the actual metal tracks. The pdks/sky130A/libs.tech/openlane/sky130_fd_sc_hd/tracks.info contains those metal informations.
- Use
grid command inside the tkon terminal to match the tracks informations:
The grids show where the routing for the local-interconnet layer can only happen, the distance of the grid lines are the required pitch of the wire. Below, we can see that the guidelines are satisfied:
-
Next, we will extract the LEF file. The LEF file contains the cell size, port definitions, and properties which aid the placer and router tool. With that, the ports definition, port class, and port use must be set first. The instructions to set these definitions via Magic are on the vsdstdcelldesign repo.
-
Next, save the mag file with a new filename save sky130_myinverter.mag. Then type lef write on the tcon terminal. It will generate a LEF file with same name as the magfile sky130_myinverter.lef. Inside that LEF file is:
Lab Part 2 [Day 4] - Plug-in the Customized Inverter Cell to OpenLane:
Inside pdks/sky130A/libs.ref/sky130_fd_sc_hd/lib/ are the liberty timing files for SKY130 PDK which contains the timing and power parameters for each cell needed in STA. It can either be slow, typical, fast with different different supply voltages (1v80, 1v65, 1v95, etc.). These are the so called PVT corners. The library name sky130_fd_sc_hd__ss_025C_1v80 describes the PVT corner as slow-slow (delay is maximum), 25° Celsius temperature, at 1.8V power supply. Timing and power parameter of a cell is obtained by simulating the cell in a variety of operating conditions (different corners) and these data are represented in the liberty file.
The liberty file characterizes all cells and is used by the ABC script during synthesis stage which maps the generic cells to the actual standard cells available in the liberty file. Same cell functionality but different sizes are available inside the library. As shown below, both are inverter cells but one with size-1 and another with size-4. Notice how size-4 inverter is simply four instances of size-1 inverter:
Provided inside the cloned vsdstdcelldesign are the liberty files containing the customized inverter cell.
- Copy the extracted lef file
sky130_myinverter.lef and the liberty files sky130*.lib from /openlane/vsdstdcelldesign/libs to the src directory of picorv32a. Open each liberty files then change the cell name sky130_vsdinv to sky130_myinverter to match the new LEF file cell name.
- Add the folowing to
config.tcl inside the picorv32a:
set ::env(LIB_SYNTH) "$::env(OPENLANE_ROOT)/designs/picorv32a/scr/sly130_fd_sc_hd__typical.lib"
set ::env(LIB_FASTEST) "$::env(OPENLANE_ROOT)/designs/picorv32a/scr/sly130_fd_sc_hd__fast.lib"
set ::env(LIB_SLOWEST) "$::env(OPENLANE_ROOT)/designs/picorv32a/scr/sly130_fd_sc_hd__slow.lib"
set ::env(LIB_TYPICAL) "$::env(OPENLANE_ROOT)/designs/picorv32a/scr/sly130_fd_sc_hd__typical.lib"
set ::env(EXTRA_LEFS) [glob $::env(OPENLANE_ROOT)/designs/$::env(DESIGN_NAME)/src/*.lef]
This sets the liberty file that will be used for ABC mapping of synthesis (LIB_SYNTH) and for STA (_FASTEST,_SLOWEST,_TYPICAL) and also the extra LEF files (EXTRA_LEFS) for the customized inverter cell. The whole config.tcl then is:
- Run docker and prepare the design picorv32a. Plug the new lef file to the OpenLANE flow via:
set lefs [glob $::env(DESIGN_DIR)/src/*.lef]
add_lefs -src $lefs
- Next
run_synthesis. Below is the synthesis statistics report runs/[date]/reports/synthesis/1-synthesis.AREA_0.stat.rpt after the run, and as we can see sky130_myinverter cell is successfully included in the design!
HOWEVER, looking at the STA log runs/[date]/logs/synthesis/sta.log, we are not meeting timing. Our neext goal is to solve this negative slack:
Delay Table:
In order to avoid large skew between endpoints of a clock tree (signal arrives at different point in time):
- Buffers on the same level must have same capacitive load to ensure same timing delay or latency on the same level.
- Buffers on the same level must also be the same size (different buffer sizes -> different W/L ratio -> different resistance -> different RC constant -> different delay).
Buffers on different level will have different capacitive load and buffer size but as long as they are the same load and size on the same level, the total delay for each clock tree path will be the same thus skew will remain zero. This means different levels will have varying input transition and output capacitive load and thus varying delay.
Delay tables are used to capture the timing model of each cell and is included inside the liberty file. The main factor in delay is the output slew. The output slew in turn depends on capacitive load and input slew. The input slew is a function of previous buffer’s output cap load and input slew and it also has its own transition delay table.
Notice how skew is zero since delay for both clock path is x9’+y15.
Lab Part 3 [Day 4] - Fix Negative Slack:
- Let us change some variables to minimize the negative slack. We will now change the variables “on the flight”. Use
echo $::env(SYNTH_STRATEGY) to view the current value of the variables before changing it:
% echo $::env(SYNTH_STRATEGY)
AREA 0
% set ::env(SYNTH_STRATEGY) "DELAY 0"
% echo $::env(SYNTH_BUFFERING)
1
% echo $::env(SYNTH_SIZING)
0
% set ::env(SYNTH_SIZING) 1
% echo $::env(SYNTH_DRIVING_CELL)
sky130_fd_sc_hd__inv_2
With SYNTH_STRATEGY of Delay 0, the tool will focus more on optimizing/minimizing the delay, index can be 0 to 3 where 3 is the most optimized for timing (sacrificing more area). SYNTH_BUFFERING of 1 ensures cell buffer will be used on high fanout cells to reduce delay due to high capacitance load. SYNTH_SIZING of 1 will enable cell sizing where cell will be upsize or downsized as needed to meet timing. SYNTH_DRIVING_CELL is the cell used to drive the input ports and is vital for cells with a lot of fan-outs since it needs higher drive strength (larger driving cell needed).
-
Below is the log report for slack and area. The area becomes bigger (from 98492 to 103364) but no negative slack anymore (from -1.2ns to +0.35ns)!
-
Next, we do run_floorplan HOWEVER:
The solution for this error is found on this issue thread. basic_macro_placement command is failing since EXTRA_LEFS variable inside config.tcl is assumed as a macro which is not. The temporary solution is to comment call on basic_macro_placement inside the OpenLane/scripts/tcl_commands/floorplan.tcl (this is okay since we are not adding any macro to the design).
-
After that run_placement, another error will occur relating to remove_buffers, the solution is to comment the call to remove_buffers_from_nets in OpenLane/scripts/tcl_commands/placement.tcl. After successfully running placement, runs/[date]/results/placement/picorv32.def will be created.
Lab Part 4 [Day 4] - Locating the Custom Inverter Cell in Layout:
-
Search for instance of cell sky130_myinverter inside the DEF file after placement stage: cat picorv32.def | grep sky130_myinverter:
-
Open the def file via magic:
magic -T /home/angelo/Desktop/OpenLane/pdks/sky130A/libs.tech/magic/sky130A.tech lef read ../../tmp/merged.nom.lef def read picorv32.def &
Select a single sky130_myinverter cell instance from the list dumped by grep (e.g. _07237_). On tkcon, command % select cell _07237_ then ctrl+z to zoom into that cell. As shown below, our customized inverter cell sky130_myinverter is sucessfully placed on the core. Use expand on tkon to show the footprint of the cell and notice how the power and ground of sky130_myinverter overlaps the power and ground pins of its adjacent cells.
Timing Analysis (Pre-Layout STA using Ideal Clocks):
Pre-layout STA will not yet include effects of clock buffers and net-delay due to RC parasitics (wire delay will be derived from PDK library wire model).
Setup timing analysis equation is:
Θ
> README 内容过长，已截取前 64 KB。完整内容请前往 GitHub 仓库查看。
评论