Skip to content

Latest commit

 

History

History
247 lines (141 loc) · 16 KB

hardening_macros.md

File metadata and controls

247 lines (141 loc) · 16 KB
Raw

Hardening Macros

Using openlane, you can produce a GDSII from an RTL for macros, and then use these macros to create your chip. Check this for more details about chip integration.

In this document we will go through the hardening steps and discuss in some detail what considerations should be made when hardening your macro.

NOTE: For all the configurations mentioned in this documentation and any other openlane configurations, you can use the exploration script run_designs.py to find the optimal value for each configuration for your design. Read more here.

Base Requirements:

You should start by setting the basic configuration file for your design. Check this for how to add your new design.

The basic configurations config.tcl should at least contain:

set ::env(DESIGN_NAME) <the name of the design>

set ::env(VERILOG_FILES) <pointer to the verilog source(s) of the design>
set ::env(CLOCK_PORT) <list of clock ports in the design rtl>
set ::env(DESIGN_IS_CORE) 0 # Since you are hardening a macro.

If you are hardening the chip core. Then make sure to set ::env(DESIGN_IS_CORE) to 1 and check this for more details about chip integration.

In case your design doesn't have a clock port, then you can either omit ::env(CLOCK_PORT) or give it an empty string.

These configurations should get you through the flow with the all other configurations using openlane default values, read about those here. However, in the coming sections we will take a closer look on how to determine the best values for most of the other configurations.

Synthesis

The first decision in synthesis is determining the optimal synthesis strategy ::env(SYNTH_STRATEGY) for your design. For that purpose there is a flag in the flow.tcl script, -synth_explore that runs a synthesis strategy exploration and reports the results in a table under <run_path>/reports/.

Then you need to consider the best values for the SYNTH_MAX_FANOUT.

If your macro is huge (200k+ cells), then you might want to try setting SYNTH_NO_FLAT to 1, which will postpone the flattening of the design during synthesis until the very end.

Other configurations like SYNTH_SIZING, SYNTH_BUFFERING, and other synthesis configurations don't have to be changed. However, the advanced user can check this documentation for more details about those configurations and their values.

Static Timing Analysis

Static Timing Analysis happens multiple times during the flow. However, they all use the same base.sdc file. You can control the outcome of the static timing analysis by setting those values:

  1. The clock ports in the design, explained in the base requirements section CLOCK_PORT.

  2. The clock period that you prefer the design to run with. This could be set using ::env(CLOCK_PERIOD) and the unit is ns. It's important to note that the flow will use this value to calculate the worst and total negative slack, also if timing optimizations are enabled, it will try to optimize for it and give suggested clock period at the end of the run in <run-path>/reports/final_summary_report.csv This value should be used in the future to speed up the optimization process and it will be the estimated value at which the design should run.

  3. The IO delay percentage from the clock period IO_PCT. More about that here.

  4. You may want to write a custom SDC file to be used in STA and CTS. The default SDC file in the flow is this file. However, you can change that by pointing to a new file with the environment variable BASE_SDC_FILE. More about that here.

Other values are set based on the (PDK, STD_CELL_LIBRARY) used. You can read more about those configurations here.

Static Timing Analysis are done after:

  1. Synthesis using the verilog netlist.

  2. Placement using OpenROAD's estimate_parasitics.

  3. Timing optimizations using the verilog netlist.

  4. Global Routing using OpenROAD's estimate_parasitics.

  5. Detailed Routing using SPEF extraction and the verilog netlist.

For SPEF extraction, you can control the wire model and the edge capacitance factor through these variables SPEF_WIRE_MODEL and SPEF_EDGE_CAP_FACTOR.

More about that here.

Floorplan

During Floor plan, you have one of three options:

  1. Let the tools determine the area relative to the size and number of cells. This is done by setting FP_SIZING to relative (the default value), and setting FP_CORE_UTIL as the core utilization percentage. Also, you can change the aspect ratio by changing FP_ASPECT_RATIO.

  2. Set a specific DIE AREA by making FP_SIZING set to absolute and then giving the size as four coordinates to DIE_AREA.

  3. Use a template DEF and apply the same DIE AREA and dimensions of that DEF. Not that this option will also force the flow to use the same PIN locations and PIN names (they are copied over from the template DEF). To use that option set: FP_DEF_TEMPLATE to point to that DEF file.

You can read more about how to control these variables here.

IO Placement

For IO placement, you currently have four options:

  1. Using a template DEF file and applying the same PIN locations and PIN names (they are copied over from the template DEF). Note that this will force the flow to apply the same exact DIE AREA and dimensions of the template DEF. To use that option set: FP_DEF_TEMPLATE to point to that DEF file.

  2. Manually setting the direction of each pin using a configuration file such as this one. Then you need to set FP_PIN_ORDER_CFG to point to that file.

  3. Using contextualized pin placement, which will automatically optimize the placement of the pins based on their context in the larger design that includes them. This relevant for macros since they will be included inside a core, and also relevant for the core since it will be part of a bigger chip. For this to happen, you need to point to the LEF and DEF of the container/parent design using these two variables: FP_CONTEXT_DEF and FP_CONTEXT_LEF. Note that currently this script can only handle the existance of a single instance of that macro.

  4. To let the tool randomly assign IOs using the random equidistant mode. This is the default way.

The options are prioritized based on the order mentioned above. This means that if you set a value for FP_DEF_TEMPLATE it will be used and the rest - if they exist - will be ignored.

You can read more about those configurations here.

Placement

Placement is done in three steps: Global Placement, Optimizations, and Detailed Placement.

Global Placement:

For Global Placement, the most important value would be PL_TARGET_DENSITY which should be easy to set.

  • If your design is not a tiny design, then PL_TARGET_DENSITY should have a value that is FP_CORE_UTIL + 1~5%. Note that FP_CORE_UTIL has a value from 0 to 100, while PL_TARGET_DENSITY has a value from 0 to 1.0.

  • If your design is a tiny design, then you may need to set PL_RANDOM_GLB_PLACEMENT to 1 or PL_RANDOM_INITIAL_PLACEMENT to 1. Also, PL_TARGET_DENSITY should have high value, while FP_CORE_UTIL should have a low value. (i.e PL_TARGET_DENSITY set to 0.5 and FP_CORE_UTIL set to 5). In very tiny designs (i.e. 1 std cell designs), the approximated DIE AREA in the floorplan stage may not leave enough room to insert tap cells in the design. Thus, it's recommended to use FP_SIZING as absolute and manually setting an appropriate DIE_AREA, check the floorplan section for more details. You may also want to reduce the values for FP_HORIZONTAL_HALO and FP_VERTICAL_HALO. You can read more about those here.

Other values to be considered are PL_BASIC_PLACEMENT and PL_SKIP_INITIAL_PLACEMENT, you can read more about those here.

Optimizations:

For this step we rely on Resizer and OpenPhySyn.

Resizer optimizations:

The only optimization we use from resizer is the wire length optimization which is used to reduce the antenna violations. This is disabled by default since the diode insertion strategies should cover that purpose.

However, you can enable that by setting PL_RESIZER_OVERBUFFER to 1 and then determine the maximum wire length by setting this value MAX_WIRE_LENGTH.

OpenPhySyn optimizations:

The timing OpenPhySyn optimizations are enabled by default and you can control that using this flag PL_OPENPHYSYN_OPTIMIZATIONS.

You can also control resizing and pin swapping using these flags: PSN_ENABLE_RESIZING and PSN_ENABLE_PIN_SWAP.

You can read more about those configurations here.

Detailed Placement:

The only value to consider here is the CELL_PAD which is usually selected for each (PDK,STD_CELL_LIBRARY) and should mostly be left as is. However, typically for the skywater libraries the value should be 4~6.

You can read more about that here.

Clock Tree Synthesis

Most of the values for clock tree synthesis are (PDK, STD_CELL_LIBRARY) specific and you can read more about those here.

You can disable it by setting CLOCK_TREE_SYNTH to 0.

If you don't want all the clock ports to be used in clock tree synthesis, then you can use set CLOCK_NET to specify those ports. Otherwise, CLOCK_NET will be defaulted to the value of CLOCK_PORT.

Power Grid (pdn)

You can skip reading this section and jump ahead to reading the advanced power grid controls to get an overall vision of how the whole chip would be powered and as such make a more educated decision in this stage. This detailed documentation exists here.

Each macro in your design should configure the common_pdn.tcl for it's use. Generally, you want to announce that the design is a macro inside the core, and that it doesn't have a core ring. Also, prohibit the router from using metal 5 by setting the maximum routing layer to met4 (layer 5). This is done by setting the following configs:

set ::env(DESIGN_IS_CORE) 0
set ::env(FP_PDN_CORE_RING) 0
set ::env(GLB_RT_MAXLAYER) 5

Pdngen configs for macros contain a special stdcell section, instead of the one used for the core in the common_pdn.tcl. The purpose of this is to prohibit the use of metal 5 in the power grid of the macros and use it exclusively for the core and top level.

If your macro contains other macros inside it. Then make sure to check the macro section and see if it requires any modifications for each them depending on their special configs. The following example is using default connect section and a different straps section:

pdngen::specify_grid macro {
    orient {R0 R180 MX MY R90 R270 MXR90 MYR90}
    power_pins "vpwr"
    ground_pins "vgnd"
    straps {
	    met1 {width 0.74 pitch 3.56 offset 0}
	    met4 {width $::env(FP_PDN_VWIDTH) pitch $::env(FP_PDN_VPITCH) offset $::env(FP_PDN_VOFFSET)}
    }
    connect {{met1 met4}}
}

WARNING: only use met1 and met4 for rails and straps.

Refer to this for more details about the syntax in case you needed to create your own pdn.tcl and then point to it using ::env(PDN_CFG).

  • The height of each macro must be greater than or equal to the value of $::env(FP_PDN_HPITCH) to allow at least two metal 5 straps on the core level to cross it and all the dropping of a via from met5 to met4 connecting the vertical straps of the macro to the horizontal straps of the core and so connect the power grid of the macro to the outer core ring.

If your design is a core, then check the power routing section in the chip integration documentation. However, the basic thing you'll need to do is to announce that your design is a chip core and that it should have a core ring. This is done by setting the following values:

set ::env(DESIGN_IS_CORE) 1
set ::env(FP_PDN_CORE_RING) 1

For more details about power grid control automation at the core and macro level and using multiple power domains please check this documentation.

Diode Insertion

Here, you have four options to choose from and they are controlled by setting DIODE_INSERTION_STRATEGY to one of the following values: (0, 1, 2, 3, 4, 5)

  1. No diode insertion is done.

  2. A diode is inserted for each PIN and connected to it.

  3. A fake diode is inserted for each PIN and connected to it, then after an antenna check is run and the fake diodes are replaced with real ones if the pin is violated.

  4. Rely on OpenROAD:FastRoute antenna avoidance flow to insert the diodes during global routing by using the Antenna Rule Checker and fixing violations. You can execute this iteratively by setting GLB_RT_MAX_DIODE_INS_ITERS, it is capable to detect any divergence, so, you'll probably end up with the lowest # of Antenna violations possible.

  5. A smarter version of strategy 1 that attempts to reduce the number of inserted diodes and places a diode at each design pin.

  6. A variant of 2 utilizing the script used in strategy 4.

You can read more about those configurations here.

Routing

The configurations here were optimized based on a large design set and are best left as is, however the advanced user could refer to this documentation to learn more about each used configuration and how to change it.

You are advised to change ROUTING_CORES based on your CPU capacity to specify the number of threads that TritonRoute can run with to perform Detailed Routing in the least runtime possible.

GDS Streaming

The configuartions here were selected based on a large design test set and the consulation of the magic sources; therefore they are best left as is. However, for the curious user, refer to this documentation to learn more about each used configuration and how to change it.

Final Reports and Checks

Finally, the flow ends with physical verification. This begins by streaming out the GDS followed by running DRC, LVS, and Antenna checks on the design. Then, it produced a final summary reports in csv format to summaries all the reports.

You can control whether the magic DRC should be done on GDS-II or on LEF/DEF abstract views. We recommend using GDS-II on macros while using LEF/DEF on the chip level. This should speed up the run process and still give as accurate as possible results. This is controlled by MAGIC_DRC_USE_GDS.

You can run Antenna Checks using OpenROAD ARC or magic. This is controlled by USE_ARC_ANTENNA_CHECK. The magic antenna checker was more reliable at the time of writing this documentation but it comes with a huge runetime trade-off and the accuracy gain is not significant enough to accept that tradeoff; thus, the default is OpenROAD's ARC.

You can control whether LVS should be run down to the device level or the cell level based on the type of the extraction, if you perform extraction on GDS-II then it's going to be down to the device/transistor level, otherwise using the LEF/DEF views then it's going to be down to the cell/block level. This is controlloed by MAGIC_EXT_USE_GDS.

You can enable LEC on the different netlists by setting LEC_ENABLE to one, which should run logic verification after writing each intermediate netlist.

A final summary report is produced by default as <run-path>/reports/final_summary_report.csv, for more details about the contents of the report check this documentation.

A final manufacturability report is produced by default as <run-path>/reports/manufacturability_report.csv, this report contains the magic DRC, the LVS, and the antenna violations summaries.

The final GDS-II should be found under <run-path>/results/magic/.

To integrate that macro into a core or a chip, check this documentation on chip integration.

If you want to create further tweaks in the flow that the abundant configurations doesn't allow, make sure to check this for more details about the interactive mode of the OpenLANE flow.