README.developers.md

Commit Procedures

For general guidance on contributing to VTR see Submitting Code to VTR.

The actual machanics of submitting code are outlined below.

However they differ slightly depending on whether you are:

• an internal developer (i.e. you have commit access to the main VTR repository at github.com/verilog-to-routing/vtr-verilog-to-routing) or,
• an (external developer) (i.e. no commit access).

The overall approach is similar, but we call out the differences below.

1. Setup a local repository on your development machine.

   a. External Developers

   • Create a 'fork' of the VTR repository.

     Usually this is done on GitHub, giving you a copy of the VTR repository (i.e. github.com/<username>/vtr-verilog-to-routing, where <username> is your GitHub username) to which you have commit rights. See About forks in the GitHub documentation.

   • Clone your 'fork' onto your local machine.

     For example, git clone [email protected]:<username>/vtr-verilog-to-routing.git, where <username> is your GitHub username.

   b. Internal Developers

   • Clone the main VTR repository onto your local machine.

     For example, git clone [email protected]:verilog-to-routing/vtr-verilog-to-routing.git.

2. Move into the cloned repository.

   For example, cd vtr-verilog-to-routing.

3. Create a branch, based off of master to work on.

   For example, git checkout -b my_awesome_branch master, where my_awesome_branch is some helpful (and descriptive) name you give you're branch. Please try to pick descriptive branch names!

4. Make your changes to the VTR code base.

5. Test your changes to ensure they work as intended and have not broken other features.

   At the bare minimum it is recommended to run:

   make                                                #Rebuild the code
   ./run_reg_test.py vtr_reg_basic vtr_reg_strong      #Run tests
   

   See Running Tests for more details.

   Also note that additional code formatting checks, and tests will be run when you open a Pull Request.

6. Commit your changes (i.e. git add followed by git commit).

   Please try to use good commit messages!

   See Commit Messages and Structure for details.

7. Push the changes to GitHub.

   For example, git push origin my_awesome_branch.

   a. External Developers

   Your code changes will now exist in your branch (e.g. my_awesome_branch) within your fork (e.g. github.com/<username>/vtr-verilog-to-routing/tree/my_awesome_branch, where <username> is your GitHub username)

   b. Internal Developers

   Your code changes will now exist in your branch (e.g. my_awesome_branch) within the main VTR repository (i.e. github.com/verilog-to-routing/vtr-verilog-to-routing/tree/my_awesome_branch)

8. Create a Pull Request (PR) to request your changes be merged into VTR.

   • Navitage to your branch on GitHub

     a. External Developers

     Navigate to your branch within your fork on GitHub (e.g. https://github.com/<username/vtr-verilog-to-routing/tree/my_awesome_branch, where <username> is your GitHub username, and my_awesome_branch is your branch name).

     b. Internal Developers

     Navigate to your branch on GitHub (e.g. https://github.com/verilog-to-routing/vtr-verilog-to-routing/tree/my_awesome_branch, where my_awesome_branch is your branch name).

   • Select the New pull request button.

     a. External Developers

     If prompted, select verilog-to-routing/vtr-verilog-to-routing as the base repository.

Commit Messages and Structure

Commit Messages

Commit messagaes are an important part of understanding the code base and it's history. It is therefore extremely important to provide the following information in the commit message:

• What is being changed?
• Why is this change occurring?

The diff of changes included with the commit provides the details of what is actually changed, so only a high-level description of what is being done is needed. However a code diff provides no insight into why the change is being made, so this extremely helpful context can only be encoded in the commit message.

The preferred convention in VTR is to structure commit messages as follows:

Header line: explain the commit in one line (use the imperative)

More detailed explanatory text. Explain the problem that this commit
is solving. Focus on why you are making this change as opposed to how
(the code explains that). Are there side effects or other unintuitive
consequences of this change? Here's the place to explain them.

If necessary. Wrap lines at some reasonable point (e.g. 74 characters,
or so) In some contexts, the header line is treated as the subject
of the commit and the rest of the text as the body. The blank line
separating the summary from the body is critical (unless you omit
the body entirely); various tools like `log`, `shortlog` and `rebase`
can get confused if you run the two together.

Further paragraphs come after blank lines.

 - Bullet points are okay, too

 - Typically a hyphen or asterisk is used for the bullet, preceded
   by a single space, with blank lines in between, but conventions
   vary here

You can also put issue tracker references at the bottom like this:

Fixes: #123
See also: #456, #789

(based off of here, and here).

Commit messages do not always need to be long, so use your judgement. More complex or involved changes with wider ranging implications likely deserve longer commit messages than fixing a simple typo.

It is often helpful to phrase the first line of a commit as an imperative/command written as if to tell the repository what to do (e.g. Update netlist data structure comments, Add tests for feature XYZ, Fix bug which ...).

To provide quick context, some VTR developers also tag the first line with the main part of the code base effected, some common ones include:

• vpr: for the VPR place and route tool (vpr/)
• flow: VTR flow architectures, scripts, tests, ... (vtr_flow/)
• archfpga: for FPGA architecture library (libs/libarchfpga)
• vtrutil: for common VTR utilities (libs/libvtrutil)
• doc: Documentation (doc/, *.md, ...)
• infra: Infrastructure (CI, .github/, ...)

Commit Structure

Generally, you should strive to keep commits atomic (i.e. they do one logical change to the code). This often means keeping commits small and focused in what they change. Of course, a large number of miniscule commits is also unhelpful (overwhelming and difficult to see the structure), and sometimes things can only be done in large changes -- so use your judgement. A reasonable rule of thumb is to try and ensure VTR will still compile after each commit.

For those familiar with history re-writing features in git (e.g. rebase) you can sometimes use these to clean-up your commit history after the fact. However these should only be done on private branches, and never directly on master.

Code Formatting

Some parts of the VTR code base (e.g. VPR, libarchfpga, libvtrutil) have C/C++ code formatting requirements which are checked automatically by regression tests. If your code changes are not compliant with the formatting, you can run:

make format

from the root of the VTR source tree. This will automatically reformat your code to be compliant with formatting requirements (this requires the clang-format tool to be available on your system).

Python code must also be compliant with the formatting. To format Python code, you can run:

make format-py

from the root of the VTR source tree (this requires the black tool to be available on your system).

Large Scale Reformatting

For large scale reformatting (should only be performed by VTR maintainers) the script dev/autoformat.py can be used to reformat the C/C++ code and commit it as 'VTR Robot', which keeps the revision history clearer and records metadata about reformatting commits (which allows git hyper-blame to skip such commits). The --python option can be used for large scale formatting of Python code.

Python Linting

Python files are automatically checked using pylint to ensure they follow established Python conventions. You can check an individual Python file by running pylint <your_python_file>, or check the entire repository by running ./dev/pylint_check.py.

Running Tests

VTR has a variety of tests which are used to check for correctness, performance and Quality of Result (QoR).

There are 4 main regression tests:

• vtr_reg_basic: ~1 minute serial

  Goal: Fast functionality check

  Feature Coverage: Low

  Benchmarks: A few small and simple circuits

  Architectures: A few simple architectures

  This regression test is not suitable for evaluating QoR or performance. It's primary purpose is to make sure the various tools do not crash/fail in the basic VTR flow.

  QoR checks in this regression test are primarily 'canary' checks to catch gross degradations in QoR. Occasionally, code changes can cause QoR failures (e.g. due to CAD noise -- particularly on small benchmarks); usually such failures are not a concern if the QoR differences are small.

• vtr_reg_strong: ~20 minutes serial, ~15 minutes with -j4

  Goal: Broad functionality check

  Feature Coverage: High

  Benchmarks: A few small circuits, with some special benchmarks to exercise specific features

  Architectures: A variety of architectures, including special architectures to exercise specific features

  This regression test is not suitable for evaluating QoR or performance. It's primary purpose is try and achieve high functionality coverage.

  QoR checks in this regression test are primarily 'canary' checks to catch gross degradations in QoR. Occasionally, changes can cause QoR failures (e.g. due to CAD noise -- particularly on small benchmarks); usually such failures are not a concern if the QoR differences are small.

• vtr_reg_nightly: ~6 hours with -j3

  Goal: Basic QoR and Performance evaluation.

  Feature Coverage: Medium

  Benchmarks: Small-medium size, diverse. Includes:

  • MCNC20 benchmarks
  • VTR benchmarks
  • Titan 'other' benchmarks (smaller than Titan23)

  Architectures: A wider variety of architectures

  QoR checks in this regression are aimed at evaluating quality and run-time of the VTR flow. As a result any QoR failures are a concern and should be investigated and understood.

• vtr_reg_weekly: ~42 hours with -j4

  Goal: Full QoR and Performance evaluation.

  Feature Coverage: Medium

  Benchmarks: Medium-Large size, diverse. Includes:

  • VTR benchmarks
  • Titan23 benchmarks

  Architectures: A wide variety of architectures

  QoR checks in this regression are aimed at evaluating quality and run-time of the VTR flow. As a result any QoR failures are a concern and should be investigated and understood.

These can be run with run_reg_test.py:

#From the VTR root directory
$ ./run_reg_test.py vtr_reg_basic
$ ./run_reg_test.py vtr_reg_strong

The nightly and weekly regressions require the Titan and ISPD benchmarks which can be integrated into your VTR tree with:

make get_titan_benchmarks
make get_ispd_benchmarks

They can then be run using run_reg_test.py:

$ ./run_reg_test.py vtr_reg_nightly
$ ./run_reg_test.py vtr_reg_weekly

To speed-up things up, individual sub-tests can be run in parallel using the -j option:

#Run up to 4 tests in parallel
$ ./run_reg_test.py vtr_reg_strong -j4

You can also run multiple regression tests together:

#Run both the basic and strong regression, with up to 4 tests in parallel
$ ./run_reg_test.py vtr_reg_basic vtr_reg_strong -j4

Odin Functionality Tests

Odin has its own set of tests to verify the correctness of its synthesis results:

• odin_reg_micro: ~2 minutes serial
• odin_reg_full: ~6 minutes serial

These can be run with:

#From the VTR root directory
$ ./run_reg_test.py odin_reg_micro
$ ./run_reg_test.py odin_reg_full

and should be used when making changes to Odin.

Unit Tests

VTR also has a limited set of unit tests, which can be run with:

#From the VTR root directory
$ make && make test

Running tests on Pull Requests (PRs) via Kokoro

Because of the long runtime for nightly and weekly tests, a Kokoro job can be used to run these tests once a Pull Request (PR) has been made at https://github.com/verilog-to-routing/vtr-verilog-to-routing.

Any pull request made by a contributor of the verilog-to-routing GitHub project on https://github.com/verilog-to-routing/ will get a set of jobs immediately. Non-contributors can request a contributor on the project add a label "kokoro:force-run" to the PR. Kokoro will then detect the tag, remove the tag, and then and issue jobs for that PR. If the tag remains after being added, there may not be an available Kokoro runner, so wait.

Re-running tests on Kokoro

If a job fails due to an intermittent failure or a re-run is desired, a contributor can add the label "kokoro:force-run" to re-issue jobs for that PR.

Checking results from Kokoro tests

Currently there is not a way for an in-flight job to be monitored.

Once a job has been completed, you can follow the "Details" link that appears on the PR status. The Kokoro page will show the job's stdout in the 'Target Log' tab (once the job has completed). The full log can be downloading by clicking the 'Download Full Log' button, or from the 'Artifacts' tab.

Downloading logs from Google Cloud Storage (GCS)

After a Kokoro run is complete a number of useful log files (e.g. for each VPR invocation) are stored to Google Cloud Storage (GCS).

The top level directory containing all VTR Kokoro runs is:

https://console.cloud.google.com/storage/browser/vtr-verilog-to-routing/artifacts/prod/foss-fpga-tools/verilog-to-routing/upstream/

PR jobs are under the presubmit directory, and continuous jobs (which run on the master branch) are under the continuous directory.

Each Kokoro run has a unique build number, which can be found in the log file (available via the Kokoro run webpage). For example, if the log file contains:

export KOKORO_BUILD_NUMBER="450"

then the Kokoro build number is 450.

If build 450 corresponded to a PR (presubmit) build of the nightly regression tests, the resulting output files would be available at:

https://console.cloud.google.com/storage/browser/vtr-verilog-to-routing/artifacts/prod/foss-fpga-tools/verilog-to-routing/upstream/presubmit/nightly/450/

where presubmit/nightly/450/ (the type, test name and build number) have been appended to the base URL. Navigating to that URL will allow you to browse and download the collected log files.

To download all the files from that Kokoro run, replace https://console.cloud.google.com/storage/browser/ in the URL with gs:// and invoke the gsutil command (and it's cp -R sub-command), like so:

gsutil -m cp -R gs://vtr-verilog-to-routing/artifacts/prod/foss-fpga-tools/verilog-to-routing/upstream/presubmit/nightly/450 .

This will download all of the logs to the current directory for inspection.

Kokoro runner details

Kokoro runners are a standard n1-highmem-16 VM with a 4 TB pd-standard disk used to perform the build of VPR and run the tests.

What to do if Kokoro jobs are not starting?

There are several reasons Kokoro jobs might not be starting. Try adding the "kokoro:force-run" label if it is not already added, or remove and add it if it already was added.

If adding the label has no affect, check GCS status, as a GCS disruption will also disrupt Kokoro.

Another reason jobs may not start is if there is a large backlog of jobs running, there may be no runners left to start. In this case, someone with Kokoro management rights may need to terminate stale jobs, or wait for job timeouts.

Debugging Failed Tests

If a test fails you probably want to look at the log files to determine the cause.

Lets assume we have a failure in vtr_reg_basic:

#In the VTR root directory
$ ./run_reg_test.py vtr_reg_strong
#Output trimmed...
regression_tests/vtr_reg_basic/basic_no_timing
-----------------------------------------
k4_N10_memSize16384_memData64/ch_intrinsics/common   failed: vpr
k4_N10_memSize16384_memData64/diffeq1/common         failed: vpr
#Output trimmed...
regression_tests/vtr_reg_basic/basic_no_timing...[Fail]
 k4_N10_memSize16384_memData64.xml/ch_intrinsics.v vpr_status: golden = success result = exited
#Output trimmed...
Error: 10 tests failed!

Here we can see that vpr failed, which caused subsequent QoR failures ([Fail]), and resulted in 10 total errors.

To see the log files we need to find the run directory. We can see from the output that the specific test which failed was regression_tests/vtr_reg_basic/basic_no_timing. All the regression tests take place under vtr_flow/tasks, so the test directory is vtr_flow/tasks/regression_tests/vtr_reg_basic/basic_no_timing. Lets move to that directory:

#From the VTR root directory
$ cd vtr_flow/tasks/regression_tests/vtr_reg_basic/basic_no_timing
$ ls
config  run001  run003
latest  run002  run004  run005

There we see there is a config directory (which defines the test), and a set of run-directories. Each time a test is run it creates a new runXXX directory (where XXX is an incrementing number). From the above we can tell that our last run was run005 (the symbolic link latest also points to the most recent run directory). From the output of run_reg_test.py we know that one of the failing architecture/circuit/parameters combinations was k4_N10_memSize16384_memData64/ch_intrinsics/common. Each architecture/circuit/parameter combination is run in its own sub-folder. Lets move to that directory:

$ cd run005/k4_N10_memSize16384_memData64/ch_intrinsics/common
$ ls
abc.out                     k4_N10_memSize16384_memData64.xml  qor_results.txt
ch_intrinsics.net           odin.out                           thread_1.out
ch_intrinsics.place         output.log                         vpr.out
ch_intrinsics.pre-vpr.blif  output.txt                         vpr_stdout.log
ch_intrinsics.route         parse_results.txt

Here we can see the individual log files produced by each tool (e.g. vpr.out), which we can use to guide our debugging. We could also manually re-run the tools (e.g. with a debugger) using files in this directory.

Evaluating Quality of Result (QoR) Changes

VTR uses highly tuned and optimized algorithms and data structures. Changes which effect these can have significant impacts on the quality of VTR's design implementations (timing, area etc.) and VTR's run-time/memory usage. Such changes need to be evaluated carefully before they are pushed/merged to ensure no quality degradation occurs.

If you are unsure of what level of QoR evaluation is necessary for your changes, please ask a VTR developer for guidance.

General QoR Evaluation Principles

The goal of performing a QoR evaluation is to measure precisely the impact of a set of code/architecture/benchmark changes on both the quality of VTR's design implementation (i.e. the result of VTR's optimizations), and on tool run-time and memory usage.

This process is made more challenging by the fact that many of VTR's optimization algorithms are based on heuristics (some of which depend on randomization). This means that VTR's implementation results are dependent upon:

• The initial conditions (e.g. input architecture & netlist, random number generator seed), and
• The precise optimization algorithms used.

The result is that a minor change to either of these can can make the measured QoR change. This effect can be viewed as an intrinsic 'noise' or 'variance' to any QoR measurement for a particular architecture/benchmark/algorithm combination.

There are typically two key methods used to measure the 'true' QoR:

1. Averaging metrics across multiple architectures and benchmark circuits.

2. Averaging metrics multiple runs of the same architecture and benchmark, but using different random number generator seeds

   This is a further variance reduction technique, although it can be very CPU-time intensive. A typical example would be to sweep an entire benchmark set across 3 or 5 different seeds.

In practice any algorithm changes will likely cause improvements on some architecture/benchmark combinations, and degradations on others. As a result we primarily focus on the average behaviour of a change to evaluate its impact. However extreme outlier behaviour on particular circuits is also important, since it may indicate bugs or other unexpected behaviour.

Key QoR Metrics

The following are key QoR metrics which should be used to evaluate the impact of changes in VTR.

Implementation Quality Metrics:

Metric	Meaning	Sensitivity
num_pre_packed_blocks	Number of primitive netlist blocks (after tech. mapping, before packing)	Low
num_post_packed_blocks	Number of Clustered Blocks (after packing)	Medium
device_grid_tiles	FPGA size in grid tiles	Low-Medium
min_chan_width	The minimum routable channel width	Medium*
crit_path_routed_wirelength	The routed wirelength at the relaxed channel width	Medium
critical_path_delay	The critical path delay at the relaxed channel width	Medium-High

* By default, VPR attempts to find the minimum routable channel width; it then performs routing at a relaxed (e.g. 1.3x minimum) channel width. At minimum channel width routing congestion can distort the true timing/wirelength characteristics. Combined with the fact that most FPGA architectures are built with an abundance of routing, post-routing metrics are usually only evaluated at the relaxed channel width.

Run-time/Memory Usage Metrics:

Metric	Meaning	Sensitivity
vtr_flow_elapsed_time	Wall-clock time to complete the VTR flow	Low
pack_time	Wall-clock time VPR spent during packing	Low
place_time	Wall-clock time VPR spent during placement	Low
min_chan_width_route_time	Wall-clock time VPR spent during routing at the minimum routable channel width	High*
crit_path_route_time	Wall-clock time VPR spent during routing at the relaxed channel width	Low
max_vpr_mem	Maximum memory used by VPR (in kilobytes)	Low

* Note that the minimum channel width route time is chaotic and can be highly variable (e.g. 10x variation is not unusual). Minimum channel width routing performs a binary search to find the minimum channel width. Since route time is highly dependent on congestion, run-time is highly dependent on the precise channel widths searched (which may change due to perturbations).

In practice you will likely want to consider additional and more detailed metrics, particularly those directly related to the changes you are making. For example, if your change related to hold-time optimization you would want to include hold-time related metrics such as hold_TNS (hold total negative slack) and hold_WNS (hold worst negative slack). If your change related to packing, you would want to report additional packing-related metrics, such as the number of clusters formed by each block type (e.g. numbers of CLBs, RAMs, DSPs, IOs).

Benchmark Selection

An important factor in performing any QoR evaluation is the benchmark set selected. In order to draw reasonably general conclusions about the impact of a change we desire two characteristics of the benchmark set:

1. It includes a large number of benchmarks which are representative of the application domains of interest.

   This ensures we don't over-tune to a specific benchmark or application domain.

2. It should include benchmarks of large sizes.

   This ensures we can optimize and scale to large problem spaces.

In practice (for various reasons) satisfying both of these goals simultaneously is challenging. The key goal here is to ensure the benchmark set is not unreasonably biased in some manner (e.g. benchmarks which are too small, benchmarks too skewed to a particular application domain).

Fairly measuring tool run-time

Accurately and fairly measuring the run-time of computer programs is challenging in practice. A variety of factors effect run-time including:

• Operating System
• System load (e.g. other programs running)
• Variance in hardware performance (e.g. different CPUs on different machines, CPU frequency scaling)

To make reasonably 'fair' run-time comparisons it is important to isolate the change as much as possible from other factors. This involves keeping as much of the experimental environment identical as possible including:

1. Target benchmarks
2. Target architecture
3. Code base (e.g. VTR revision)
4. CAD parameters
5. Computer system (e.g. CPU model, CPU frequency/power scaling, OS version)
6. Compiler version

Collecting QoR Measurements

The first step is to collect QoR metrics on your selected benchmark set.

You need at least two sets of QoR measurements:

1. The baseline QoR (i.e. unmodified VTR).
2. The modified QoR (i.e. VTR with your changes).

Note that it is important to generate both sets of QoR measurements on the same computing infrastructure to ensure a fair run-time comparison.

The following examples show how a single set of QoR measurements can be produced using the VTR flow infrastructure.

Example: VTR Benchmarks QoR Measurement

The VTR benchmarks are a group of benchmark circuits distributed with the VTR project. The are provided as synthesizable verilog and can be re-mapped to VTR supported architectures. They consist mostly of small to medium sized circuits from a mix of application domains. They are used primarily to evaluate the VTR's optimization quality in an architecture exploration/evaluation setting (e.g. determining minimum channel widths).

A typical approach to evaluating an algorithm change would be to run vtr_reg_qor_chain task from the nightly regression test:

#From the VTR root
$ cd vtr_flow/tasks

#Run the VTR benchmarks
$ ../scripts/run_vtr_task.py regression_tests/vtr_reg_nightly/vtr_reg_qor_chain

#Several hours later... they complete

#Parse the results
$ ../scripts/python_libs/vtr/parse_vtr_task.py regression_tests/vtr_reg_nightly/vtr_reg_qor_chain

#The run directory should now contain a summary parse_results.txt file
$ head -5 vtr_reg_nightly/vtr_reg_qor_chain/latest/parse_results.txt
arch                                  	circuit           	script_params	vpr_revision 	vpr_status	error	num_pre_packed_nets	num_pre_packed_blocks	num_post_packed_nets	num_post_packed_blocks	device_width	device_height	num_clb	num_io	num_outputs	num_memoriesnum_mult	placed_wirelength_est	placed_CPD_est	placed_setup_TNS_est	placed_setup_WNS_est	min_chan_width	routed_wirelength	min_chan_width_route_success_iteration	crit_path_routed_wirelength	crit_path_route_success_iteration	critical_path_delay	setup_TNS	setup_WNS	hold_TNS	hold_WNS	logic_block_area_total	logic_block_area_used	min_chan_width_routing_area_total	min_chan_width_routing_area_per_tile	crit_path_routing_area_total	crit_path_routing_area_per_tile	odin_synth_time	abc_synth_time	abc_cec_time	abc_sec_time	ace_time	pack_time	place_time	min_chan_width_route_time	crit_path_route_time	vtr_flow_elapsed_time	max_vpr_mem	max_odin_mem	max_abc_mem
k6_frac_N10_frac_chain_mem32K_40nm.xml	bgm.v             	common       	9f591f6-dirty	success   	     	26431              	24575                	14738               	2258                  	53          	53           	1958   	257   	32         	0           11      	871090               	18.5121       	-13652.6            	-18.5121            	84            	328781           	32                                    	297718                     	18                               	20.4406            	-15027.8 	-20.4406 	0       	0       	1.70873e+08           	1.09883e+08          	1.63166e+07                      	5595.54                             	2.07456e+07                 	7114.41                        	11.16          	1.03          	-1          	-1          	-1      	141.53   	108.26    	142.42                   	15.63               	652.17               	1329712    	528868      	146796
k6_frac_N10_frac_chain_mem32K_40nm.xml	blob_merge.v      	common       	9f591f6-dirty	success   	     	14163              	11407                	3445                	700                   	30          	30           	564    	36    	100        	0           0       	113369               	13.4111       	-2338.12            	-13.4111            	64            	80075            	18                                    	75615                      	23                               	15.3479            	-2659.17 	-15.3479 	0       	0       	4.8774e+07            	3.03962e+07          	3.87092e+06                      	4301.02                             	4.83441e+06                 	5371.56                        	0.46           	0.17          	-1          	-1          	-1      	67.89    	11.30     	47.60                    	3.48                	198.58               	307756     	48148       	58104
k6_frac_N10_frac_chain_mem32K_40nm.xml	boundtop.v        	common       	9f591f6-dirty	success   	     	1071               	1141                 	595                 	389                   	13          	13           	55     	142   	192        	0           0       	5360                 	3.2524        	-466.039            	-3.2524             	34            	4534             	15                                    	3767                       	12                               	3.96224            	-559.389 	-3.96224 	0       	0       	6.63067e+06           	2.96417e+06          	353000.                          	2088.76                             	434699.                     	2572.18                        	0.29           	0.11          	-1          	-1          	-1      	2.55     	0.82      	2.10                     	0.15                	7.24                 	87552      	38484       	37384
k6_frac_N10_frac_chain_mem32K_40nm.xml	ch_intrinsics.v   	common       	9f591f6-dirty	success   	     	363                	493                  	270                 	247                   	10          	10           	17     	99    	130        	1           0       	1792                 	1.86527       	-194.602            	-1.86527            	46            	1562             	13                                    	1438                       	20                               	2.4542             	-226.033 	-2.4542  	0       	0       	3.92691e+06           	1.4642e+06           	259806.                          	2598.06                             	333135.                     	3331.35                        	0.03           	0.01          	-1          	-1          	-1      	0.46     	0.31      	0.94                     	0.09                	2.59                 	62684      	8672        	32940

Example: Titan Benchmarks QoR Measurements

The Titan benchmarks are a group of large benchmark circuits from a wide range of applications, which are compatible with the VTR project. The are typically used as post-technology mapped netlists which have been pre-synthesized with Quartus. They are substantially larger and more realistic than the VTR benchmarks, but can only target specifically compatible architectures. They are used primarily to evaluate the optimization quality and scalability of VTR's CAD algorithms while targeting a fixed architecture (e.g. at a fixed channel width).

A typical approach to evaluating an algorithm change would be to run vtr_reg_titan task from the weekly regression test:

#From the VTR root

#Download and integrate the Titan benchmarks into the VTR source tree
$ make get_titan_benchmarks

#Move to the task directory
$ cd vtr_flow/tasks

#Run the VTR benchmarks
$ ../scripts/run_vtr_task.py regression_tests/vtr_reg_weekly/vtr_reg_titan

#Several days later... they complete

#Parse the results
$ ../scripts/python_libs/vtr/parse_vtr_task.py regression_tests/vtr_reg_weekly/vtr_reg_titan

#The run directory should now contain a summary parse_results.txt file
$ head -5 vtr_reg_nightly/vtr_reg_qor_chain/latest/parse_results.txt
arch                     	circuit                                 	vpr_revision	vpr_status	error	num_pre_packed_nets	num_pre_packed_blocks	num_post_packed_nets	num_post_packed_blocks	device_width	device_height	num_clb	num_io	num_outputs	num_memoriesnum_mult	placed_wirelength_est	placed_CPD_est	placed_setup_TNS_est	placed_setup_WNS_est	routed_wirelength	crit_path_route_success_iteration	logic_block_area_total	logic_block_area_used	routing_area_total	routing_area_per_tile	critical_path_delay	setup_TNS   setup_WNS	hold_TNS	hold_WNS	pack_time	place_time	crit_path_route_time	max_vpr_mem	max_odin_mem	max_abc_mem
stratixiv_arch.timing.xml	neuron_stratixiv_arch_timing.blif       	0208312     	success   	     	119888             	86875                	51408               	3370                  	128         	95           	-1     	42    	35         	-1          -1      	3985635              	8.70971       	-234032             	-8.70971            	1086419          	20                               	0                     	0                    	2.66512e+08       	21917.1              	9.64877            	-262034     -9.64877 	0       	0       	127.92   	218.48    	259.96              	5133800    	-1          	-1
stratixiv_arch.timing.xml	sparcT1_core_stratixiv_arch_timing.blif 	0208312     	success   	     	92813              	91974                	54564               	4170                  	77          	57           	-1     	173   	137        	-1          -1      	3213593              	7.87734       	-534295             	-7.87734            	1527941          	43                               	0                     	0                    	9.64428e+07       	21973.8              	9.06977            	-625483     -9.06977 	0       	0       	327.38   	338.65    	364.46              	3690032    	-1          	-1
stratixiv_arch.timing.xml	stereo_vision_stratixiv_arch_timing.blif	0208312     	success   	     	127088             	94088                	62912               	3776                  	128         	95           	-1     	326   	681        	-1          -1      	4875541              	8.77339       	-166097             	-8.77339            	998408           	16                               	0                     	0                    	2.66512e+08       	21917.1              	9.36528            	-187552     -9.36528 	0       	0       	110.03   	214.16    	189.83              	5048580    	-1          	-1
stratixiv_arch.timing.xml	cholesky_mc_stratixiv_arch_timing.blif  	0208312     	success   	     	140214             	108592               	67410               	5444                  	121         	90           	-1     	111   	151        	-1          -1      	5221059              	8.16972       	-454610             	-8.16972            	1518597          	15                               	0                     	0                    	2.38657e+08       	21915.3              	9.34704            	-531231     -9.34704 	0       	0       	211.12   	364.32    	490.24              	6356252    	-1          	-1

Comparing QoR Measurements

Once you have two (or more) sets of QoR measurements they now need to be compared.

A general method is as follows:

1. Normalize all metrics to the values in the baseline measurements (this makes the relative changes easy to evaluate)
2. Produce tables for each set of QoR measurements showing the per-benchmark relative values for each metric
3. Calculate the GEOMEAN over all benchmarks for each normalized metric
4. Produce a summary table showing the Metric Geomeans for each set of QoR measurements

QoR Comparison Gotchas

There are a variety of 'gotchas' you need to avoid to ensure fair comparisons:

• GEOMEAN's must be over the same set of benchmarks . A common issue is that a benchmark failed to complete for some reason, and it's metric values are missing

• Run-times need to be collected on the same compute infrastructure at the same system load (ideally unloaded).

Example QoR Comparison

Suppose we've make a change to VTR, and we now want to evaluate the change. As described above we produce QoR measurements for both the VTR baseline, and our modified version.

We then have the following (hypothetical) QoR Metrics.

Baseline QoR Metrics:

arch	circuit	num_pre_packed_blocks	num_post_packed_blocks	device_grid_tiles	min_chan_width	crit_path_routed_wirelength	critical_path_delay	vtr_flow_elapsed_time	pack_time	place_time	min_chan_width_route_time	crit_path_route_time	max_vpr_mem
k6_frac_N10_frac_chain_mem32K_40nm.xml	bgm.v	24575	2258	2809	84	297718	20.4406	652.17	141.53	108.26	142.42	15.63	1329712
k6_frac_N10_frac_chain_mem32K_40nm.xml	blob_merge.v	11407	700	900	64	75615	15.3479	198.58	67.89	11.3	47.6	3.48	307756
k6_frac_N10_frac_chain_mem32K_40nm.xml	boundtop.v	1141	389	169	34	3767	3.96224	7.24	2.55	0.82	2.1	0.15	87552
k6_frac_N10_frac_chain_mem32K_40nm.xml	ch_intrinsics.v	493	247	100	46	1438	2.4542	2.59	0.46	0.31	0.94	0.09	62684
k6_frac_N10_frac_chain_mem32K_40nm.xml	diffeq1.v	886	313	256	60	9624	17.9648	15.59	2.45	1.36	9.93	0.93	86524
k6_frac_N10_frac_chain_mem32K_40nm.xml	diffeq2.v	599	201	256	52	8928	13.7083	13.14	1.41	0.87	9.14	0.94	85760
k6_frac_N10_frac_chain_mem32K_40nm.xml	LU8PEEng.v	31396	2286	2916	100	348085	79.4512	1514.51	175.67	153.01	1009.08	45.47	1410872
k6_frac_N10_frac_chain_mem32K_40nm.xml	LU32PEEng.v	101542	7251	9216	158	1554942	80.062	28051.68	625.03	930.58	25050.73	251.87	4647936
k6_frac_N10_frac_chain_mem32K_40nm.xml	mcml.v	165809	6767	8649	128	1311825	51.1905	9088.1	524.8	742.85	4001.03	127.42	4999124
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkDelayWorker32B.v	4145	1327	2500	38	30086	8.39902	65.54	7.73	15.39	26.19	3.23	804720
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkPktMerge.v	1160	516	784	44	13370	4.4408	21.75	2.45	2.14	13.95	1.96	122872
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkSMAdapter4B.v	2852	548	400	48	19274	5.26765	47.64	16.22	4.16	19.95	1.14	116012
k6_frac_N10_frac_chain_mem32K_40nm.xml	or1200.v	4530	1321	729	62	51633	9.67406	105.62	33.37	12.93	44.95	3.33	219376
k6_frac_N10_frac_chain_mem32K_40nm.xml	raygentop.v	2934	710	361	58	22045	5.14713	39.72	9.54	4.06	19.8	2.34	126056
k6_frac_N10_frac_chain_mem32K_40nm.xml	sha.v	3024	236	289	62	16653	10.0144	390.89	11.47	2.7	6.18	0.75	117612
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision0.v	21801	1122	1156	58	64935	3.63177	82.74	20.45	15.49	24.5	2.6	411884
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision1.v	19538	1096	1600	100	143517	5.61925	272.41	26.99	18.15	149.46	15.49	676844
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision2.v	42078	2534	7396	134	650583	15.3151	3664.98	66.72	119.26	3388.7	62.6	3114880
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision3.v	324	55	49	30	768	2.66429	2.25	0.75	0.2	0.57	0.05	61148

Modified QoR Metrics:

arch	circuit	num_pre_packed_blocks	num_post_packed_blocks	device_grid_tiles	min_chan_width	crit_path_routed_wirelength	critical_path_delay	vtr_flow_elapsed_time	pack_time	place_time	min_chan_width_route_time	crit_path_route_time	max_vpr_mem
k6_frac_N10_frac_chain_mem32K_40nm.xml	bgm.v	24575	2193	2809	82	303891	20.414	642.01	70.09	113.58	198.09	16.27	1222072
k6_frac_N10_frac_chain_mem32K_40nm.xml	blob_merge.v	11407	684	900	72	77261	14.6676	178.16	34.31	13.38	57.89	3.35	281468
k6_frac_N10_frac_chain_mem32K_40nm.xml	boundtop.v	1141	369	169	40	3465	3.5255	4.48	1.13	0.7	0.9	0.17	82912
k6_frac_N10_frac_chain_mem32K_40nm.xml	ch_intrinsics.v	493	241	100	54	1424	2.50601	1.75	0.19	0.27	0.43	0.09	60796
k6_frac_N10_frac_chain_mem32K_40nm.xml	diffeq1.v	886	293	256	50	9972	17.3124	15.24	0.69	0.97	11.27	1.44	72204
k6_frac_N10_frac_chain_mem32K_40nm.xml	diffeq2.v	599	187	256	50	7621	13.1714	14.14	0.63	1.04	10.93	0.78	68900
k6_frac_N10_frac_chain_mem32K_40nm.xml	LU8PEEng.v	31396	2236	2916	98	349074	77.8611	1269.26	88.44	153.25	843.31	49.13	1319276
k6_frac_N10_frac_chain_mem32K_40nm.xml	LU32PEEng.v	101542	6933	9216	176	1700697	80.1368	28290.01	306.21	897.95	25668.4	278.74	4224048
k6_frac_N10_frac_chain_mem32K_40nm.xml	mcml.v	165809	6435	8649	124	1240060	45.6693	9384.4	296.99	686.27	4782.43	99.4	4370788
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkDelayWorker32B.v	4145	1207	2500	36	33354	8.3986	53.94	3.85	14.75	19.53	2.95	785316
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkPktMerge.v	1160	494	784	36	13881	4.57189	20.75	0.82	1.97	15.01	1.88	117636
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkSMAdapter4B.v	2852	529	400	56	19817	5.21349	27.58	5.05	2.66	14.65	1.11	103060
k6_frac_N10_frac_chain_mem32K_40nm.xml	or1200.v	4530	1008	729	76	48034	8.70797	202.25	10.1	8.31	171.96	2.86	178712
k6_frac_N10_frac_chain_mem32K_40nm.xml	raygentop.v	2934	634	361	58	20799	5.04571	22.58	2.75	2.42	12.86	1.64	108116
k6_frac_N10_frac_chain_mem32K_40nm.xml	sha.v	3024	236	289	62	16052	10.5007	337.19	5.32	2.25	4.52	0.69	105948
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision0.v	21801	1121	1156	58	70046	3.61684	86.5	9.5	15.02	41.81	2.59	376100
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision1.v	19538	1080	1600	92	142805	6.02319	343.83	10.68	16.21	247.99	11.66	480352
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision2.v	42078	2416	7396	124	646793	14.6606	5614.79	34.81	107.66	5383.58	62.27	2682976
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision3.v	324	54	49	34	920	2.5281	1.55	0.31	0.14	0.43	0.05	63444

Based on these metrics we then calculate the following ratios and summary.

QoR Metric Ratio (Modified QoR / Baseline QoR):

arch	circuit	num_pre_packed_blocks	num_post_packed_blocks	device_grid_tiles	min_chan_width	crit_path_routed_wirelength	critical_path_delay	vtr_flow_elapsed_time	pack_time	place_time	min_chan_width_route_time	crit_path_route_time	max_vpr_mem
k6_frac_N10_frac_chain_mem32K_40nm.xml	bgm.v	1.00	0.97	1.00	0.98	1.02	1.00	0.98	0.50	1.05	1.39	1.04	0.92
k6_frac_N10_frac_chain_mem32K_40nm.xml	blob_merge.v	1.00	0.98	1.00	1.13	1.02	0.96	0.90	0.51	1.18	1.22	0.96	0.91
k6_frac_N10_frac_chain_mem32K_40nm.xml	boundtop.v	1.00	0.95	1.00	1.18	0.92	0.89	0.62	0.44	0.85	0.43	1.13	0.95
k6_frac_N10_frac_chain_mem32K_40nm.xml	ch_intrinsics.v	1.00	0.98	1.00	1.17	0.99	1.02	0.68	0.41	0.87	0.46	1.00	0.97
k6_frac_N10_frac_chain_mem32K_40nm.xml	diffeq1.v	1.00	0.94	1.00	0.83	1.04	0.96	0.98	0.28	0.71	1.13	1.55	0.83
k6_frac_N10_frac_chain_mem32K_40nm.xml	diffeq2.v	1.00	0.93	1.00	0.96	0.85	0.96	1.08	0.45	1.20	1.20	0.83	0.80
k6_frac_N10_frac_chain_mem32K_40nm.xml	LU8PEEng.v	1.00	0.98	1.00	0.98	1.00	0.98	0.84	0.50	1.00	0.84	1.08	0.94
k6_frac_N10_frac_chain_mem32K_40nm.xml	LU32PEEng.v	1.00	0.96	1.00	1.11	1.09	1.00	1.01	0.49	0.96	1.02	1.11	0.91
k6_frac_N10_frac_chain_mem32K_40nm.xml	mcml.v	1.00	0.95	1.00	0.97	0.95	0.89	1.03	0.57	0.92	1.20	0.78	0.87
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkDelayWorker32B.v	1.00	0.91	1.00	0.95	1.11	1.00	0.82	0.50	0.96	0.75	0.91	0.98
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkPktMerge.v	1.00	0.96	1.00	0.82	1.04	1.03	0.95	0.33	0.92	1.08	0.96	0.96
k6_frac_N10_frac_chain_mem32K_40nm.xml	mkSMAdapter4B.v	1.00	0.97	1.00	1.17	1.03	0.99	0.58	0.31	0.64	0.73	0.97	0.89
k6_frac_N10_frac_chain_mem32K_40nm.xml	or1200.v	1.00	0.76	1.00	1.23	0.93	0.90	1.91	0.30	0.64	3.83	0.86	0.81
k6_frac_N10_frac_chain_mem32K_40nm.xml	raygentop.v	1.00	0.89	1.00	1.00	0.94	0.98	0.57	0.29	0.60	0.65	0.70	0.86
k6_frac_N10_frac_chain_mem32K_40nm.xml	sha.v	1.00	1.00	1.00	1.00	0.96	1.05	0.86	0.46	0.83	0.73	0.92	0.90
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision0.v	1.00	1.00	1.00	1.00	1.08	1.00	1.05	0.46	0.97	1.71	1.00	0.91
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision1.v	1.00	0.99	1.00	0.92	1.00	1.07	1.26	0.40	0.89	1.66	0.75	0.71
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision2.v	1.00	0.95	1.00	0.93	0.99	0.96	1.53	0.52	0.90	1.59	0.99	0.86
k6_frac_N10_frac_chain_mem32K_40nm.xml	stereovision3.v	1.00	0.98	1.00	1.13	1.20	0.95	0.69	0.41	0.70	0.75	1.00	1.04
	GEOMEAN	1.00	0.95	1.00	1.02	1.01	0.98	0.92	0.42	0.87	1.03	0.96	0.89

QoR Summary:

	baseline	modified
num_pre_packed_blocks	1.00	1.00
num_post_packed_blocks	1.00	0.95
device_grid_tiles	1.00	1.00
min_chan_width	1.00	1.02
crit_path_routed_wirelength	1.00	1.01
critical_path_delay	1.00	0.98
vtr_flow_elapsed_time	1.00	0.92
pack_time	1.00	0.42
place_time	1.00	0.87
min_chan_width_route_time	1.00	1.03
crit_path_route_time	1.00	0.96
max_vpr_mem	1.00	0.89

From the results we can see that our change, on average, achieved a small reduction in the number of logic blocks (0.95) in return for a 2% increase in minimum channel width and 1% increase in routed wirelength. From a run-time perspective the packer is substantially faster (0.42).

Automated QoR Comparison Script

To automate some of the QoR comparison VTR includes a script to compare parse_results.txt files and generate a spreadsheet including the ratio and summary tables.

For example:

#From the VTR Root
$ ./vtr_flow/scripts/qor_compare.py parse_results1.txt parse_results2.txt parse_results3.txt -o comparison.xlsx

will produce ratio tables and a summary table for the files parse_results1.txt, parse_results2.txt and parse_results3.txt, where the first file (parse_results1.txt) is assumed to be the baseline used to produce normalized ratios.

Generating New QoR Golden Result

There may be times when a regression test fails its QoR test because its golden_result needs to be changed due to known changes in code behaviour. In this case, a new golden result needs to be generated so that the test can be passed. To generate a new golden result, follow the steps outlined below.

1. Move to the vtr_flow/tasks directory from the VTR root, and run the failing test. For example, if a test called vtr_ex_test in vtr_reg_nightly was failing:

   #From the VTR root
   $ cd vtr_flow/tasks
   $ ../scripts/run_vtr_task.py regression_tests/vtr_reg_nightly/vtr_ex_test

2. Next, generate new golden reference results using parse_vtr_task.py and the -create_golden option.

   $ ../scripts/parse_vtr_task.py regression_tests/vtr_reg_nightly/vtr_ex_test -create_golden

3. Lastly, check that the results match with the -check_golden option

   $ ../scripts/parse_vtr_task.py regression_tests/vtr_reg_nightly/vtr_ex_test -check_golden

Once the -check_golden command passes, the changes to the golden result can be committed so that the reg test will pass in future runs of vtr_reg_nightly.

Adding Tests

Any time you add a feature to VTR you must add a test which exercises the feature. This ensures that regression tests will detect if the feature breaks in the future.

Consider which regression test suite your test should be added to (see Running Tests descriptions).

Typically, test which exercise new features should be added to vtr_reg_strong. These tests should use small benchmarks to ensure they:

• run quickly (so they get run often!), and
• are easier to debug. If your test will take more than ~1 minute it should probably go in a longer running regression test (but see first if you can create a smaller testcase first).

Adding a test to vtr_reg_strong

This describes adding a test to vtr_reg_strong, but the process is similar for the other regression tests.

1. Create a configuration file

   First move to the vtr_reg_strong directory:

   #From the VTR root directory
   $ cd vtr_flow/tasks/regression_tests/vtr_reg_strong
   $ ls
   qor_geomean.txt             strong_flyover_wires        strong_pack_and_place
   strong_analysis_only        strong_fpu_hard_block_arch  strong_power
   strong_bounding_box         strong_fracturable_luts     strong_route_only
   strong_breadth_first        strong_func_formal_flow     strong_scale_delay_budgets
   strong_constant_outputs     strong_func_formal_vpr      strong_sweep_constant_outputs
   strong_custom_grid          strong_global_routing       strong_timing
   strong_custom_pin_locs      strong_manual_annealing     strong_titan
   strong_custom_switch_block  strong_mcnc                 strong_valgrind
   strong_echo_files           strong_minimax_budgets      strong_verify_rr_graph
   strong_fc_abs               strong_multiclock           task_list.txt
   strong_fix_pins_pad_file    strong_no_timing            task_summary
   strong_fix_pins_random      strong_pack

   Each folder (prefixed with strong_ in this case) defines a task (sub-test).

   Let's make a new task named strong_mytest. An easy way is to copy an existing configuration file such as strong_timing/config/config.txt

   $ mkdir -p strong_mytest/config
   $ cp strong_timing/config/config.txt strong_mytest/config/.

   You can now edit strong_mytest/config/config.txt to customize your test.

2. Generate golden reference results

   Now we need to test our new test and generate 'golden' reference results. These will be used to compare future runs of our test to detect any changes in behaviour (e.g. bugs).

   From the VTR root, we move to the vtr_flow/tasks directory, and then run our new test:

   #From the VTR root
   $ cd vtr_flow/tasks
   $ ../scripts/run_vtr_task.py regression_tests/vtr_reg_strong/strong_mytest
   
   regression_tests/vtr_reg_strong/strong_mytest
   -----------------------------------------
   Current time: Jan-25 06:51 PM.  Expected runtime of next benchmark: Unknown
   k6_frac_N10_mem32K_40nm/ch_intrinsics...OK

   Next we can generate the golden reference results using parse_vtr_task.py with the -create_golden option:

   $ ../scripts/python_libs/vtr/parse_vtr_task.py regression_tests/vtr_reg_strong/strong_mytest -create_golden

   And check that everything matches with -check_golden:

   $ ../scripts/python_libs/vtr/parse_vtr_task.py regression_tests/vtr_reg_strong/strong_mytest -check_golden
   regression_tests/vtr_reg_strong/strong_mytest...[Pass]

3. Add it to the task list

   We now need to add our new strong_mytest task to the task list, so it is run whenever vtr_reg_strong is run. We do this by adding the line regression_tests/vtr_reg_strong/strong_mytest to the end of vtr_reg_strong's task_list.txt:

   #From the VTR root directory
   $ vim vtr_flow/tasks/regression_tests/vtr_reg_strong/task_list.txt
   # Add a new line 'regression_tests/vtr_reg_strong/strong_mytest' to the end of the file

   Now, when we run vtr_reg_strong:

   #From the VTR root directory
   $ ./run_reg_test.py vtr_reg_strong
   #Output trimmed...
   regression_tests/vtr_reg_strong/strong_mytest
   -----------------------------------------
   #Output trimmed...

   we see our test is run.

4. Commit the new test

   Finally you need to commit your test:

   #Add the config.txt and golden_results.txt for the test
   $ git add vtr_flow/tasks/regression_tests/vtr_reg_strong/strong_mytest/
   #Add the change to the task_list.txt
   $ git add vtr_flow/tasks/regression_tests/vtr_reg_strong/task_list.txt
   #Commit the changes, when pushed the test will automatically be picked up by BuildBot
   $ git commit

Debugging Aids

VTR has support for several additional tools/features to aid debugging.

Sanitizers

VTR can be compiled using sanitizers which will detect invalid memory accesses, memory leaks and undefined behaviour (supported by both GCC and LLVM):

#From the VTR root directory
$ cmake -D VTR_ENABLE_SANITIZE=ON build
$ make

Assertion Levels

VTR supports configurable assertion levels.

The default level (2) which turns on most assertions which don't cause significant run-time penalties.

This level can be increased:

#From the VTR root directory
$ cmake -D VTR_ASSERT_LEVEL=3 build
$ make

this turns on more extensive assertion checking and re-builds VTR.

GDB Pretty Printers

To make it easier to debug some of VTR's data structures with GDB.

STL Pretty Printers

It is helpful to enable STL pretty printers, which make it much easier to debug data structures using STL.

For example printing a std::vector<int> by default prints:

(gdb) p/r x_locs
$2 = {<std::_Vector_base<int, std::allocator<int> >> = {
    _M_impl = {<std::allocator<int>> = {<__gnu_cxx::new_allocator<int>> = {<No data fields>}, <No data fields>}, _M_start = 0x555556f063b0, 
      _M_finish = 0x555556f063dc, _M_end_of_storage = 0x555556f064b0}}, <No data fields>}

which is not very helpful.

But with STL pretty printers it prints:

(gdb) p x_locs
$2 = std::vector of length 11, capacity 64 = {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10}

which is much more helpful for debugging!

If STL pretty printers aren't already enabled on your system, add the following to your .gdbinit file:

python
import sys
sys.path.insert(0, '$STL_PRINTER_ROOT')
from libstdcxx.v6.printers import register_libstdcxx_printers
register_libstdcxx_printers(None)

end

where $STL_PRINTER_ROOT should be replaced with the appropriate path to the STL pretty printers. For example recent versions of GCC include these under /usr/share/gcc-*/python (e.g. /usr/share/gcc-9/python)

VTR Pretty Printers

VTR includes some pretty printers for some VPR/VTR specific types.

For example, without the pretty printers you would see the following when printing a VPR AtomBlockId:

(gdb) p blk_id
$1 = {
  id_ = 71
}

But with the VTR pretty printers enabled you would see:

(gdb) p blk_id
$1 = AtomBlockId(71)

To enable the VTR pretty printers in GDB add the following to your .gdbinit file:

python
import sys

sys.path.insert(0, "$VTR_ROOT/dev")
import vtr_gdb_pretty_printers
gdb.pretty_printers.append(vtr_gdb_pretty_printers.vtr_type_lookup)

end

where $VTR_ROOT should be replaced with the root of the VTR source tree on your system.

RR (Record Replay) Debugger

RR extends GDB with the ability to to record a run of a tool and then re-run it to reproduce any observed issues. RR also enables efficient reverse execution (!) which can be extremely helpful when tracking down the source of a bug.

Speeding up the edit-compile-test cycle

Rapid iteration through the edit-compile-test/debug cycle is very helpful when making code changes to VTR.

The following is some guidance on techniques to reduce the time required.

Speeding Compilation

1. Parallel compilation

   For instance when building VTR using make, you can specify the -j N option to compile the code base with N parallel jobs:

   $ make -j N
   

   A reasonable value for N is equal to the number of threads you system can run. For instance, if your system has 4 cores with HyperThreading (i.e. 2-way SMT) you could run:

   $ make -j8
   

2. Building only a subset of VTR

   If you know your changes only effect a specific tool in VTR, you can request that only that tool is rebuilt. For instance, if you only wanted to re-compile VPR you could run:

   $ make vpr
   

   which would avoid re-building other tools (e.g. ODIN, ABC).

3. Use ccache

   ccache is a program which caches previous compilation results. This can save significant time, for instance, when switching back and forth between release and debug builds.

   VTR's cmake configuration should automatically detect and make use of ccache once it is installed.

   For instance on Ubuntu/Debian systems you can install ccache with:

   $ sudo apt install ccache
   

   This only needs to be done once on your development system.

4. Disable Interprocedural Optimizatiaons (IPO)

   IPO re-optimizes an entire executable at link time, and is automatically enabled by VTR if a supporting compiler is found. This can notably improve performance (e.g. ~10-20% faster), but can significantly increase compilation time (e.g. >2x in some cases). When frequently re-compiling and debugging the extra execution speed may not be worth the longer compilation times. In such cases you can manually disable IPO by setting the cmake parameter VTR_IPO_BUILD=off.

   For instance using the wrapper Makefile:

   $ make CMAKE_PARAMS="-DVTR_IPO_BUILD=off"
   

   Note that this option is sticky, so subsequent calls to make don't need to keep specifying VTR_IPO_BUILD, until you want to re-enable it.

   This setting can also be changed with the ccmake tool (i.e. ccmake build).

All of these option can be used in combination. For example, the following will re-build only VPR using 8 parallel jobs with IPO disabled:

make CMAKE_PARAMS="-DVTR_IPO_BUILD=off" -j8 vpr

Profiling VTR

1. Install gprof, gprof2dot, and xdot. Specifically, the previous two packages require python3, and you should install the last one with sudo apt install for all the dependencies you will need for visualizing your profile results.

   pip3 install gprof
   pip3 install gprof2dot
   sudo apt install xdot
   

   Contact your administrator if you do not have the sudo rights.

2. Use the CMake option below to enable VPR profiler build.

   make CMAKE_PARAMS="-DVTR_ENABLE_PROFILING=ON" vpr
   

3. With the profiler build, each time you run the VTR flow script, it will produce an extra file gmon.out that contains the raw profile information. Run gprof to parse this file. You will need to specify the path to the VPR executable.

   gprof $VTR_ROOT/vpr/vpr gmon.out > gprof.txt
   

4. Next, use gprof2dot to transform the parsed results to a .dot file, which describes the graph of your final profile results. If you encounter long function names, specify the -s option for a cleaner graph.

   gprof2dot -s gprof.txt > vpr.dot
   

5. You can chain the above commands to directly produce the .dot file:

   gprof $VTR_ROOT/vpr/vpr gmon.out | gprof2dot -s > vpr.dot
   

6. Use xdot to view your results:

   xdot vpr.dot
   

7. To save your results as a png file:

   dot -Tpng -Gdpi=300 vpr.dot > vpr.png
   

   Note that you can use the -Gdpi option to make your picture clearer if you find the default dpi settings not clear enough.

External Subtrees

VTR includes some code which is developed in external repositories, and is integrated into the VTR source tree using git subtrees.

To simplify the process of working with subtrees we use the dev/external_subtrees.py script.

For instance, running ./dev/external_subtrees.py --list from the VTR root it shows the subtrees:

Component: abc             Path: abc                            URL: https://github.com/berkeley-abc/abc.git       URL_Ref: master
Component: libargparse     Path: libs/EXTERNAL/libargparse      URL: https://github.com/kmurray/libargparse.git    URL_Ref: master
Component: libblifparse    Path: libs/EXTERNAL/libblifparse     URL: https://github.com/kmurray/libblifparse.git   URL_Ref: master
Component: libsdcparse     Path: libs/EXTERNAL/libsdcparse      URL: https://github.com/kmurray/libsdcparse.git    URL_Ref: master
Component: libtatum        Path: libs/EXTERNAL/libtatum         URL: https://github.com/kmurray/tatum.git          URL_Ref: master

Code included in VTR by subtrees should not be modified within the VTR source tree. Instead changes should be made in the relevant up-stream repository, and then synced into the VTR tree.

Updating an existing Subtree

1. From the VTR root run: ./dev/external_subtrees.py $SUBTREE_NAME, where $SUBTREE_NAME is the name of an existing subtree.

   For example to update the libtatum subtree:

   ./dev/external_subtrees.py --update libtatum

Adding a new Subtree

To add a new external subtree to VTR do the following:

1. Add the subtree specification to dev/subtree_config.xml.

   For example to add a subtree name libfoo from the master branch of https://github.com/kmurray/libfoo.git to libs/EXTERNAL/libfoo you would add:

   <subtree
       name="libfoo"
       internal_path="libs/EXTERNAL/libfoo"
       external_url="https://github.com/kmurray/libfoo.git"
       default_external_ref="master"/>

   within the existing <subtrees> tag.

   Note that the internal_path directory should not already exist.

   You can confirm it works by running: dev/external_subtrees.py --list:

   Component: abc             Path: abc                            URL: https://github.com/berkeley-abc/abc.git       URL_Ref: master
   Component: libargparse     Path: libs/EXTERNAL/libargparse      URL: https://github.com/kmurray/libargparse.git    URL_Ref: master
   Component: libblifparse    Path: libs/EXTERNAL/libblifparse     URL: https://github.com/kmurray/libblifparse.git   URL_Ref: master
   Component: libsdcparse     Path: libs/EXTERNAL/libsdcparse      URL: https://github.com/kmurray/libsdcparse.git    URL_Ref: master
   Component: libtatum        Path: libs/EXTERNAL/libtatum         URL: https://github.com/kmurray/tatum.git          URL_Ref: master
   Component: libfoo          Path: libs/EXTERNAL/libfoo           URL: https://github.com/kmurray/libfoo.git         URL_Ref: master
   

   which shows libfoo is now recognized.

2. Run ./dev/external_subtrees.py --update $SUBTREE_NAME to add the subtree.

   For the libfoo example above this would be:

   ./dev/external_subtrees.py --update libfoo

   This will create two commits to the repository. The first will squash all the upstream changes, the second will merge those changes into the current branch.

Subtree Rational

VTR uses subtrees to allow easy tracking of upstream dependencies.

Their main advantages included:

• Works out-of-the-box: no actions needed post checkout to pull in dependencies (e.g. no git submodule update --init --recursive)
• Simplified upstream version tracking
• Potential for local changes (although in VTR we do not use this to make keeping in sync easier)

See here for a more detailed discussion.

Finding Bugs with Coverity

Coverity Scan is a static code analysis service which can be used to detect bugs.

Browsing Defects

To view defects detected do the following:

1. Get a coverity scan account

   Contact a project maintainer for an invitation.

2. Browse the existing defects through the coverity web interface

Submitting a build

To submit a build to coverity do the following:

1. Download the coverity build tool

2. Configure VTR to perform a debug build. This ensures that all assertions are enabled, without assertions coverity may report bugs that are guarded against by assertions. We also set VTR asserts to the highest level.

   #From the VTR root
   mkdir -p build
   cd build
   CC=gcc CXX=g++ cmake -DCMAKE_BUILD_TYPE=debug -DVTR_ASSERT_LEVEL=3 ..

Note that we explicitly asked for gcc and g++, the coverity build tool defaults to these compilers, and may not like the default 'cc' or 'c++' (even if they are linked to gcc/g++).

3. Run the coverity build tool

   #From the build directory where we ran cmake
   cov-build --dir cov-int make -j8

4. Archive the output directory

   tar -czvf vtr_coverity.tar.gz cov-int

5. Submit the archive through the coverity web interface

Once the build has been analyzed you can browse the latest results through the coverity web interface

No files emitted

If you get the following warning from cov-build:

[WARNING] No files were emitted.

You may need to configure coverity to 'know' about your compiler. For example:

```shell
cov-configure --compiler `which gcc-7`
```

On unix-like systems run scan-build make from the root VTR directory. to output the html analysis to a specific folder, run scan-build make -o /some/folder

Release Procedures

General Principles

We periodically make 'official' VTR releases. While we aim to keep the VTR master branch stable through-out development some users prefer to work of off an official release. Historically this has coincided with the publishing of a paper detailing and carefully evaluating the changes from the previous VTR release. This is particularly helpful for giving academics a named baseline version of VTR to which they can compare which has a known quality.

In preparation for a release it may make sense to produce 'release candidates' which when fully tested and evaluated (and after any bug fixes) become the official release.

Checklist

The following outlines the procedure to following when making an official VTR release:

• Check the code compiles on the list of supported compilers
• Check that all regression tests pass functionality
• Update regression test golden results to match the released version
• Check that all regression tests pass QoR
• Create a new entry in the CHANGELOG.md for the release, summarizing at a high-level user-facing changes
• Increment the version number (set in root CMakeLists.txt)
• Create a git annotated tag (e.g. v8.0.0) and push it to github
• GitHub will automatically create a release based on the tag
• Add the new change log entry to the GitHub release description
• Update the ReadTheDocs configuration to build and serve documentation for the relevant tag (e.g. v8.0.0)
• Send a release announcement email to the vtr-announce mailing list (make sure to thank all contributors!)