[lit] - Allow 1 test to report multiple micro-test results to provide support for microbenchmarks.

[homerdin]

Summary

These changes are to allow to a Result object to have nested Result objects in order to support microbenchmarks. Currently lit is restricted to reporting one result object for one test, this change provides support for when a test that wants to report individual timings for individual kernels.

This is revision is result of the discussions in https://reviews.llvm.org/D32272#794759, https://reviews.llvm.org/D37421#f8003b27 and https://reviews.llvm.org/D38496 that I have seen. It is a separation of the changes purposed in https://reviews.llvm.org/D40077.

With this change I will be adding the LCALS (Livermore Compiler Analysis Loop Suite) collection of loop kernels to the llvm test suite using the google benchmark library (https://reviews.llvm.org/D43319).

Previously microbenchmarks had been handled by using macros to section groups of microbenchmarks together and build many executables while still getting a grouped timing (MultiSource/TSVC). Recently the google benchmark library was added to the test suite and utilized with a litsupport plugin. However the limitation of 1 test 1 result limited its use to passing a runtime option to run only 1 microbenchmark with several hand written tests (MicroBenchmarks/XRay). This runs the same executable many times with different hand written tests. I will update the litsupport plugin to utilize the new functionality (https://reviews.llvm.org/D43316).

These changes allow lit to report micro test results if desired in order to get many precise timing results from 1 run of 1 test executable.

Example Output from LCALS:

Terminal:

PASS: test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test (1 of 1)
********** TEST 'test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test' RESULTS **********
MicroBenchmarks: 12 
compile_time: 6.9610 
hash: "5075a3ae907cf9631cdc4cf8401cbfb3" 
link_time: 0.0426 
**********
*** MICRO-TEST: BM_IF_QUAD_RAW/171
    exec_time:  2.6995 
*** MICRO-TEST: BM_IF_QUAD_RAW/44217
    exec_time:  698.8880 
*** MICRO-TEST: BM_IF_QUAD_RAW/5001
    exec_time:  78.9838 
*** MICRO-TEST: BM_INIT3_RAW/171
    exec_time:  0.2248 
*** MICRO-TEST: BM_INIT3_RAW/44217
    exec_time:  168.0970 
*** MICRO-TEST: BM_INIT3_RAW/5001
    exec_time:  15.1119 
*** MICRO-TEST: BM_MULADDSUB_RAW/171
    exec_time:  0.4491 
*** MICRO-TEST: BM_MULADDSUB_RAW/44217
    exec_time:  169.6760 
*** MICRO-TEST: BM_MULADDSUB_RAW/5001
    exec_time:  16.1443 
*** MICRO-TEST: BM_TRAP_INT_RAW/171
    exec_time:  2.0922 
*** MICRO-TEST: BM_TRAP_INT_RAW/44217
    exec_time:  540.9620 
*** MICRO-TEST: BM_TRAP_INT_RAW/5001
    exec_time:  61.1846

Partial JSON output:

"tests": [
  {
    "code": "PASS",
    "elapsed": null,
    "metrics": {
      "exec_time": 540.962
    },
    "name": "test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test:BM_TRAP_INT_RAW/44217",
    "output": ""
  },
  {
    "code": "PASS",
    "elapsed": null,
    "metrics": {
      "exec_time": 2.6995
    },
    "name": "test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test:BM_IF_QUAD_RAW/171",
    "output": ""
  },
  {
    "code": "PASS",
    "elapsed": null,
    "metrics": {
      "exec_time": 169.676
    },
    "name": "test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test:BM_MULADDSUB_RAW/44217",
    "output": ""
  },
  {
    "code": "PASS",
    "elapsed": null,
    "metrics": {
      "exec_time": 168.097
    },
    "name": "test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test:BM_INIT3_RAW/44217",
    "output": ""
  },
  {
    "code": "PASS",
    "elapsed": null,
    "metrics": {
      "exec_time": 16.1443
    },
    "name": "test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test:BM_MULADDSUB_RAW/5001",
    "output": ""
  },
  {
    "code": "PASS",
    "elapsed": null,
    "metrics": {
      "exec_time": 61.1846
    },
    "name": "test-suite :: MicroBenchmarks/LCALS/SubsetBRawLoops/lcalsBRaw.test:BM_TRAP_INT_RAW/5001",
    "output": ""
  },

[delcypher]

Please add at least one test to lit's test suite for the functionality you're adding. As this patch stands right now it's not obvious how to use this new functionality. A test would demonstrate how this feature is intended to be used and will ensure that if someone breaks your new feature that it will be noticed. Documentation would be great but that's probably asking too much.

[rengolin]

I think this patch should be merged with D43316 as a first example, and then used in D43319 as a proper example. Once D43316 is approved, D43319 should wait until a few green builds are out on the LNT bots to go, once approved.

Detecting LNT errors in the bots is not a trivial task sometimes and pushing only one behavioural change at a time helps a lot.

I've also added a few comments on the Python code.

Finally, please use diff -U9999 to show the context before uploading the diff. This is hard to review as it is.

Thanks,
--renato

utils/lit/lit/main.py

86

IIRC Python 2.x and 3.x behave different regarding .items().

Do we force Python 3 for LIT?

89

We already have the **** enclosure for the whole test, I think this is a bit too verbose.

I'd suggest something like *** MICRO-TEST: micro_test_name

91

guard with if micro_test.metrics:

93

Same as above, we don't need more *** :)

130

What's with the [:-5] ?

[homerdin]

I updated to a full context diff and have made some of the python changes requested. I changed the way the name is formed to make it clearer.

@delcypher I had misread your request for tests that show what I am trying to do, which is why I opened the revision for LCALS. I will add a test to lit's test suite for these changes. I'll also look into the documentation and see what I can contribute there.

I still need to look into the Python 2.x and 3.x behavior regarding .items() and if we force Python 3 for LIT.

[homerdin]

utils/lit/lit/main.py

130

I changed the way I build the name to be clearer. Was using [:-5] (.test) to expand the parent test name with the micro test name before ".test"

[homerdin]

utils/lit/lit/main.py

130

Realizing this is incorrect as it only works with ".test" files. Need a safer way to expand the test name.

[rengolin]

utils/lit/lit/main.py

130

regular expression would do nicely

[homerdin]

I added tests for the console and JSON output of micro-tests. Also made a change to the way the test name is expanded.

[rengolin]

utils/lit/tests/test-data-micro.py

4

Join these two lines with a pipe, otherwise .out files will be kept and can break further tests.

14

Both micro-test and micro-results are being stored in dictionaries, that means their order is not guaranteed.

I assume that's the reason for the {{[0-2]}} regex here, but a similar solution would be needed for the values as well.

I recommend you look at CHECK-DAG:

1. CHECK-NEXT: *** MICRO-TEST: test{{[0-2]}}
2. CHECK-DAG: micro_value0: 4 ​# CHECK-DAG: micro_value1: 1.3

This means there are two lines after MICRO-TEST which have those two values, in any order.

It could match to a list after the next MICRO-TEST, however, if you have three CHECK-NEXT: of MICRO-TEST, then the last one would fail if you match the CHECK-DAG on the wrong line, so you're covered.

Just repeat that pattern three times and it should be fine.

utils/lit/tests/test-output-micro.py

1

Same thing here, join the lines with a pipe.

9

Here, the CHECK-DAG trick won't work, because you have two levels: one for the result and another for the metrics, and there's no way to identify CHECK-DAG1 and CHECK-DAG2.

One way to solve this is to sort the keys before dumping JSON, as to get a reproducible answer (that is, of course, if Python's default sort is stable).

[homerdin]

Thanks for all the feedback! I've made some changes to the tests:

• Updated test-data-micro test to use CHECK-DAG and piped output
• Updated test-output-micro test to use CHECK-DAG. I think this should work since parent result test-data-micro :: micro-tests.ini will always be appended to the end of the list after all micro-tests. This is testing the lit's output file so I sent the stdout to /dev/null and explicitly removed the output file to clean the test up.

[MatzeB]

Looks nearly good to me. I'd just opt for a different naming of the sub-tests. Also some nitpicks/comments on other nitpicks below:

utils/lit/lit/main.py

86

.items() use should be fine here (it'll give you a full copy in python2 and a lazy iterator in python3. But except for higher memory they should behave the same, and this doesn't look like we care about the memory usage of a few entries).

134

Appending the subtest name before .test feels odd to me as there is no actual file with the name testname/microtest.test how about testname.test:microtest instead?

utils/lit/tests/test-data-micro.py

14

Actually I would recommend changing the printing code above to something like:

sorted_results = sorted(test.result.microResults.items())
for key, micro_test in sorted_results:

to avoid the indeterministic output.

utils/lit/tests/test-output-micro.py

9

lit does the json dumping with sort_keys=True so the output should be deterministic so we don't need the CHECK-DAGs.

[homerdin]

Thanks for the comments!

• Updated micro test name expansion to be ParentTest.test:MicroTest
• Updated Console output to be deterministic and removed the CHECK-DAGS from the tests
• Added Updated Output in the Summary

[rengolin]

Thanks @MatzeB, apart from the missing pipe in the RUN line, I'm happy when you are.

cheers,
--renato

utils/lit/tests/test-data-micro.py

14

Indeed, better.

utils/lit/tests/test-output-micro.py

1

Forgot the pipe here. :)

9

Ah, perfect!

[homerdin]

utils/lit/tests/test-output-micro.py

1

This test is checking the output file that lit writes, is there a way to pipe the output file to FileCheck?

[rengolin]

utils/lit/tests/test-output-micro.py

1

Yup. Try --output - for stdout, or just omit --output and see if it goes automatically to stdout.

[homerdin]

Tried some things and was able to get it working using /dev/stdout.

[rengolin]

In D43314#1021294, @homerdin wrote:

Tried some things and was able to get it working using /dev/stdout.

That doesn't work on Windows, may not work on Mac?

[homerdin]

In D43314#1025162, @rengolin wrote:

In D43314#1021294, @homerdin wrote:

Tried some things and was able to get it working using /dev/stdout.

That doesn't work on Windows, may not work on Mac?

Should I change it back to rm %t.results.out after FileCheck? I'm not sure how to do this with a pipe.

[MatzeB]

In D43314#1031716, @homerdin wrote:

In D43314#1025162, @rengolin wrote:

In D43314#1021294, @homerdin wrote:

Tried some things and was able to get it working using /dev/stdout.

That doesn't work on Windows, may not work on Mac?

Should I change it back to rm %t.results.out after FileCheck? I'm not sure how to do this with a pipe.

Yes, given that lit has nothing like -o - writing a temporary file is fine.

And I think most tests don't clean the temporary files as part of the test. I guess it's better for debugging in case of a failure if the file is still around and as the temp files are part of the build directory they will be cleaned up when the build directory is removed as a whole.

[rengolin]

I think it should be fine if it's not in the list of patterns recognised as tests...

[homerdin]

I believe I have made all the requested changes. Thank you for all the feedback!

[rengolin]

LGTM with the sorting comment. Thanks!

utils/lit/lit/main.py

130

Don't you need to sort these ones, too?

[homerdin]

utils/lit/lit/main.py

130

Both the micro_test_data and micro_metrics_data dictionaries are sorted when dumped into the json file and are not used elsewhere.

[lebedev.ri]

llvm/trunk/utils/lit/lit/Test.py

163–164 ↗

(On Diff #138213)

Was this restriction motivated?
Does LNT fail to import multiple results with the same name from a single .json?

[homerdin]

llvm/trunk/utils/lit/lit/Test.py

163–164 ↗

(On Diff #138213)

LNT accepts multiple results with the same name in a .json from multiple runs --exec-multisample as a combination of the individual lit outputs. If we give two tests the same name in the test suite we would be combining results from two separate tests. If you want to get multiple samples you can use --exec-multisample. Also google benchmark is already reporting the mean of many runs.

If you are wondering if you can add a micro-test result with the same name as a micro-test result that runs under a different executable that works because the names are extended with the parent's.

[lebedev.ri]

llvm/trunk/utils/lit/lit/Test.py

163–164 ↗

(On Diff #138213)

Thanks.
I was actually thinking about using benchmark's repetitions instead of --exec-multisample,
but it appears that doesn't work, can't pass a list of measurements into addMicroResult().

utils/lit/lit/Test.py

def __init__(self, code, output='', elapsed=None):
         # The result code.
         self.code = code
         # The test output.
         self.output = output
         # The wall timing to execute the test, if timing.
         self.elapsed = elapsed
         # The metrics reported by this test.
         self.metrics = {}
+        # The micro-test results reported by this test.
+        self.microResults = {}
 
     def addMetric(self, name, value):
         """
         addMetric(name, value)
 
         Attach a test metric to the test result, with the given name and list of
         values. It is an error to attempt to attach the metrics with the same
         name multiple times.
 
         Each value must be an instance of a MetricValue subclass.
         """
         if name in self.metrics:
             raise ValueError("result already includes metrics for %r" % (
                     name,))
         if not isinstance(value, MetricValue):
             raise TypeError("unexpected metric value: %r" % (value,))
         self.metrics[name] = value
 
+    def addMicroResult(self, name, microResult):
+        """
+        addMicroResult(microResult)
+
+        Attach a micro-test result to the test result, with the given name and
+        result.  It is an error to attempt to attach a micro-test with the 
+        same name multiple times.
+
+        Each micro-test result must be an instance of the Result class.
+        """
+        if name in self.microResults:
+            raise ValueError("Result already includes microResult for %r" % (
+                   name,))
+        if not isinstance(microResult, Result):
+            raise TypeError("unexpected MicroResult value %r" % (microResult,))
+        self.microResults[name] = microResult
+
+
 # Test classes.
 
 class TestSuite:
     """TestSuite - Information on a group of tests.
 
     A test suite groups together a set of logically related tests.
     """
 

utils/lit/lit/main.py

def update(self, test):
         if test.result.metrics:
             print("%s TEST '%s' RESULTS %s" % ('*'*10, test.getFullName(),
                                                '*'*10))
             items = sorted(test.result.metrics.items())
             for metric_name, value in items:
                 print('%s: %s ' % (metric_name, value.format()))
             print("*" * 10)
 
+        # Report micro-tests, if present
+        if test.result.microResults:
+            items = sorted(test.result.microResults.items())
+            for micro_test_name, micro_test in items:
+                print("%s MICRO-TEST: %s" %
+                         ('*'*3, micro_test_name))
+   
+                if micro_test.metrics:
+                    for metric_name, value in micro_test.metrics.items():
+                        print('    %s:  %s ' % (metric_name, value.format()))
+
         # Ensure the output is flushed.
         sys.stdout.flush()
 
 def write_test_results(run, lit_config, testing_time, output_path):
     try:
         import json
     except ImportError:
         lit_config.fatal('test output unsupported with Python 2.5')
for test in run.tests:
             'elapsed' : test.result.elapsed }
 
         # Add test metrics, if present.
         if test.result.metrics:
             test_data['metrics'] = metrics_data = {}
             for key, value in test.result.metrics.items():
                 metrics_data[key] = value.todata()
 
+        # Report micro-tests separately, if present
+        if test.result.microResults:
+            for key, micro_test in test.result.microResults.items():
+                # Expand parent test name with micro test name
+                parent_name = test.getFullName()
+                micro_test_name = '/' + key
+                # Add micro test name before test extension
+                rex = re.compile("(.+?)(\.[^.]*$|$)")
+                result = rex.match(parent_name)
+                if result:
+                    micro_full_name = result.group(1) \
+                                    + micro_test_name \
+                                    + result.group(2)
+                else:
+                    micro_full_name = parent_name + micro_test_name
+
+                micro_test_data = {
+                    'name' : micro_full_name,
+                    'code' : micro_test.code.name,
+                    'output' : micro_test.output,
+                    'elapsed' : micro_test.elapsed }
+                if micro_test.metrics:
+                    micro_test_data['metrics'] = micro_metrics_data = {}
+                    for key, value in micro_test.metrics.items():
+                        micro_metrics_data[key] = value.todata()
+
+                tests_data.append(micro_test_data)
+
         tests_data.append(test_data)
 
     # Write the output.
     f = open(output_path, 'w')
     try:
         json.dump(data, f, indent=2, sort_keys=True)
         f.write('\n')
     finally:

utils/lit/tests/Inputs/test-data-micro/dummy_format.py

+import os
+try:
+    import ConfigParser
+except ImportError:
+    import configparser as ConfigParser
+
+import lit.formats
+import lit.Test
+
+class DummyFormat(lit.formats.FileBasedTest):
+    def execute(self, test, lit_config):
+        # In this dummy format, expect that each test file is actually just a
+        # .ini format dump of the results to report.
+
+        source_path = test.getSourcePath()
+
+        cfg = ConfigParser.ConfigParser()
+        cfg.read(source_path)
+
+        # Create the basic test result.
+        result_code = cfg.get('global', 'result_code')
+        result_output = cfg.get('global', 'result_output')
+        result = lit.Test.Result(getattr(lit.Test, result_code),
+                                 result_output)
+
+        # Load additional metrics.
+        for key,value_str in cfg.items('results'):
+            value = eval(value_str)
+            if isinstance(value, int):
+                metric = lit.Test.IntMetricValue(value)
+            elif isinstance(value, float):
+                metric = lit.Test.RealMetricValue(value)
+            else:
+                raise RuntimeError("unsupported result type")
+            result.addMetric(key, metric)
+
+        # Create micro test results
+        for key,micro_name in cfg.items('micro-tests'):
+            micro_result = lit.Test.Result(getattr(lit.Test, result_code, ''))
+            # Load micro test additional metrics
+            for key,value_str in cfg.items('micro-results'):
+                value = eval(value_str)
+                if isinstance(value, int):
+                    metric = lit.Test.IntMetricValue(value)
+                elif isinstance(value, float):
+                    metric = lit.Test.RealMetricValue(value)
+                else:
+                    raise RuntimeError("unsupported result type")
+                micro_result.addMetric(key, metric)
+            result.addMicroResult(micro_name, micro_result)
+
+        return result

utils/lit/tests/Inputs/test-data-micro/lit.cfg

+import site
+site.addsitedir(os.path.dirname(__file__))
+import dummy_format
+
+config.name = 'test-data-micro'
+config.suffixes = ['.ini']
+config.test_format = dummy_format.DummyFormat()
+config.test_source_root = None
+config.test_exec_root = None
+config.target_triple = None

utils/lit/tests/Inputs/test-data-micro/micro-tests.ini

+[global]
+result_code = PASS
+result_output = Test passed.
+
+[results]
+value0 = 1
+value1 = 2.3456
+
+[micro-tests]
+microtest0 = test0
+microtest1 = test1
+microtest2 = test2
+
+[micro-results]
+micro_value0 = 4
+micro_value1 = 1.3

utils/lit/tests/test-data-micro.py

+# Test features related to formats which support reporting additional test data.
+# and multiple test results.
+
+# RUN: %{lit} -j 1 -v %{inputs}/test-data-micro > %t.out
+# RUN: FileCheck < %t.out %s
+
+# CHECK: -- Testing:
+
+# CHECK: PASS: test-data-micro :: micro-tests.ini
+# CHECK-NEXT: *** TEST 'test-data-micro :: micro-tests.ini' RESULTS ***
+# CHECK-NEXT: value0: 1
+# CHECK-NEXT: value1: 2.3456
+# CHECK-NEXT: ***
+# CHECK-NEXT: *** MICRO-TEST: test{{[0-2]}}
+# CHECK-NEXT: micro_value0: 4
+# CHECK-NEXT: micro_value1: 1.3
+# CHECK-NEXT: *** MICRO-TEST: test{{[0-2]}}
+# CHECK-NEXT: micro_value0: 4
+# CHECK-NEXT: micro_value1: 1.3
+# CHECK-NEXT: *** MICRO-TEST: test{{[0-2]}}
+# CHECK-NEXT: micro_value0: 4
+# CHECK-NEXT: micro_value1: 1.3

utils/lit/tests/test-output-micro.py

+# RUN: %{lit} -j 1 -v %{inputs}/test-data-micro --output %t.results.out > %t.out
+# RUN: FileCheck < %t.results.out %s
+
+# CHECK: {
+# CHECK: "__version__"
+# CHECK: "elapsed"
+# CHECK-NEXT: "tests": [
+# CHECK-NEXT:   {
+# CHECK-NEXT:     "code": "PASS",
+# CHECK-NEXT:     "elapsed": null,
+# CHECK-NEXT:     "metrics": {
+# CHECK-NEXT:       "micro_value0": 4,
+# CHECK-NEXT:       "micro_value1": 1.3
+# CHECK-NEXT:     }
+# CHECK-NEXT:     "name": "test-data-micro :: micro-tests/test{{[0-2]}}.ini",
+# CHECK-NEXT:     "output": ""
+# CHECK-NEXT:   }
+# CHECK-NEXT:   {
+# CHECK-NEXT:     "code": "PASS",
+# CHECK-NEXT:     "elapsed": null,
+# CHECK-NEXT:     "metrics": {
+# CHECK-NEXT:       "micro_value0": 4,
+# CHECK-NEXT:       "micro_value1": 1.3
+# CHECK-NEXT:     }
+# CHECK-NEXT:     "name": "test-data-micro :: micro-tests/test{{[0-2]}}.ini",
+# CHECK-NEXT:     "output": ""
+# CHECK-NEXT:   }
+# CHECK-NEXT:   {
+# CHECK-NEXT:     "code": "PASS",
+# CHECK-NEXT:     "elapsed": null,
+# CHECK-NEXT:     "metrics": {
+# CHECK-NEXT:       "micro_value0": 4,
+# CHECK-NEXT:       "micro_value1": 1.3
+# CHECK-NEXT:     }
+# CHECK-NEXT:     "name": "test-data-micro :: micro-tests/test{{[0-2]}}.ini",
+# CHECK-NEXT:     "output": ""
+# CHECK-NEXT:   }
+# CHECK-NEXT:   {
+# CHECK-NEXT:     "code": "PASS",
+# CHECK-NEXT:     "elapsed": {{[0-9.]+}},
+# CHECK-NEXT:     "metrics": {
+# CHECK-NEXT:       "value0": 1,
+# CHECK-NEXT:       "value1": 2.3456
+# CHECK-NEXT:     }
+# CHECK-NEXT:     "name": "test-data-micro :: micro-tests.ini",
+# CHECK-NEXT:     "output": "Test passed."
+# CHECK-NEXT:   }
+# CHECK-NEXT: ]
+# CHECK-NEXT: }