Merge pull request #123 from antmicro/jboc/benchmark

Add plotting of benchmark results

test/run_benchmarks.py

@@ -5,9 +5,12 @@
 
 import os
 import re
-import yaml
+import sys
 import argparse
 import subprocess
+from collections import defaultdict, namedtuple
+
+import yaml
 
 from litedram.common import Settings
 
@@ -18,21 +21,27 @@ from .benchmark import LiteDRAMBenchmarkSoC
 def ng(name, regex):
     return r'(?P<{}>{})'.format(name, regex)
 
+def center(text, width, fillc=' '):
+    added = width - len(text)
+    left = added // 2
+    right = added - left
+    return fillc * left + text + fillc * right
 
 def human_readable(value):
     binary_prefixes = ['', 'k', 'M', 'G', 'T']
+    mult = 1.0
     for prefix in binary_prefixes:
-        if value < 1024:
+        if value * mult < 1024:
             break
-        value /= 1024
+        mult /= 1024
-    return value, prefix
+    return mult, prefix
 
 # Benchmark configuration --------------------------------------------------------------------------
 
 class BenchmarkConfiguration(Settings):
     def __init__(self, sdram_module, sdram_data_width, bist_length, bist_random):
         self.set_attributes(locals())
-        self._settings = {k: v for k, v in locals().items() if v != self}
+        self._settings = {k: v for k, v in locals().items() if k != 'self'}
 
     def as_args(self):
         args = []
@@ -45,6 +54,12 @@ class BenchmarkConfiguration(Settings):
                 args.extend([arg_string, str(value)])
         return args
 
+    def __eq__(self, other):
+        if not isinstance(other, BenchmarkConfiguration):
+            return NotImplemented
+        return all((getattr(self, setting) == getattr(other, setting)
+                    for setting in self._settings.keys()))
+
     @classmethod
     def load_yaml(cls, yaml_file):
         with open(yaml_file) as f:
@@ -101,61 +116,166 @@ class BenchmarkResult:
 # Results summary ----------------------------------------------------------------------------------
 
 class ResultsSummary:
+    # value_scaling is a function: value -> (multiplier, prefix)
+    Fmt = namedtuple('MetricFormatting', ['name', 'unit', 'value_scaling'])
+    metric_formats = {
+        'write_bandwidth':  Fmt('Write bandwidth',  'bps', lambda value: human_readable(value)),
+        'read_bandwidth':   Fmt('Read bandwidth',   'bps', lambda value: human_readable(value)),
+        'write_efficiency': Fmt('Write efficiency', '',    lambda value: (100, '%')),
+        'read_efficiency':  Fmt('Read efficiency',  '',    lambda value: (100, '%')),
+    }
+
     def __init__(self, results):
         self.results = results
-        # convert results, which map config->metrics to a mapping metric->(config->result)
-        self.write_bandwidth = self.collect('write_bandwidth')
-        self.read_bandwidth = self.collect('read_bandwidth')
-        self.write_efficiency = self.collect('write_efficiency')
-        self.read_efficiency = self.collect('read_efficiency')
 
-    def create_name(self, config):
+    def by_metric(self, metric):
-        return '{}:{}:{}:{}'.format(
+        """Returns pairs of value of the given metric and the configuration used for benchmark"""
-            config.sdram_module, config.sdram_data_width,
-            config.bist_length, config.bist_random)
-
-    def collect(self, attribute):
-        by_case = {}
         for result in self.results:
-            value = getattr(result, attribute)()
+            value = getattr(result, metric)()
-            by_case[self.create_name(result.config)] = value
+            yield value, result.config
-        return by_case
-
-    def value_string(self, metric, value):
-        if metric in ['write_bandwidth', 'read_bandwidth']:
-            return '{:6.3f} {}bps'.format(*human_readable(value))
-        elif ['write_efficiency', 'read_efficiency']:
-            return '{:5.1f} %'.format(100 * value)
-        else:
-            raise ValueError()
 
     def print(self):
-        print('\n---====== Summary ======---')
+        legend = '(module, datawidth, length, random, result)'
-        for metric in ['write_bandwidth', 'read_bandwidth', 'write_efficiency', 'read_efficiency']:
+        fmt = '   {module:15}  {dwidth:2}  {length:4}  {random:1}    {result}'
-            print(metric)
-            for case, value in getattr(self, metric).items():
-                print('  {:30}  {}'.format(case, self.value_string(metric, value)))
 
-    def plot(self):
+        # store formatted lines per metric
-        raise NotImplementedError()
+        metric_lines = defaultdict(list)
+        for metric, (_, unit, formatter) in self.metric_formats.items():
+            for value, config in self.by_metric(metric):
+                mult, prefix = formatter(value)
+                result = '{:5.1f} {}{}'.format(value * mult, prefix, unit)
+                line = fmt.format(module=config.sdram_module,
+                                  dwidth=config.sdram_data_width,
+                                  length=config.bist_length,
+                                  random=int(config.bist_random),
+                                  result=result)
+                metric_lines[metric].append(line)
+
+        # find length of the longest line
+        max_length = max((len(l) for lines in metric_lines.values() for l in lines))
+        max_length = max(max_length, len(legend) + 2)
+        width = max_length + 2
+
+        # print the formatted summary
+        def header(text):
+            mid = center(text, width - 6, '=')
+            return center(mid, width, '-')
+        print(header(' Summary '))
+        print(center(legend, width))
+        for metric, lines in metric_lines.items():
+            print(center(self.metric_formats[metric].name, width))
+            for line in lines:
+                print(line)
+        print(header(''))
+
+    def plot(self, output_dir, backend='Agg', theme='default', save_format='png', **savefig_kwargs):
+        """Create plots with benchmark results summary
+
+        Default backend is Agg, which is non-GUI backed and only allows
+        to save figures as files. If a GUI backed is passed, plt.show()
+        will be called at the end.
+        """
+        # import locally here to be able to run benchmarks without installing matplotlib
+        import matplotlib
+        matplotlib.use(backend)
+
+        import matplotlib.pyplot as plt
+        import numpy as np
+        from matplotlib.ticker import FuncFormatter, PercentFormatter
+
+        plt.style.use(theme)
+
+        def bandwidth_formatter_func(value, pos):
+            mult, prefix = human_readable(value)
+            return '{:.1f}{}bps'.format(value * mult, prefix)
+
+        tick_formatters = {
+            'write_bandwidth':  FuncFormatter(bandwidth_formatter_func),
+            'read_bandwidth':   FuncFormatter(bandwidth_formatter_func),
+            'write_efficiency': PercentFormatter(1.0),
+            'read_efficiency':  PercentFormatter(1.0),
+        }
+
+        def config_tick_name(config):
+            return '{}\n{}, {}, {}'.format(config.sdram_module, config.sdram_data_width,
+                                         config.bist_length, int(config.bist_random))
+
+        for metric, (name, unit, _) in self.metric_formats.items():
+            fig = plt.figure()
+            axis = plt.gca()
+
+            values, configs = zip(*self.by_metric(metric))
+            ticks = np.arange(len(configs))
+
+            axis.barh(ticks, values, align='center')
+            axis.set_yticks(ticks)
+            axis.set_yticklabels([config_tick_name(c) for c in configs])
+            axis.invert_yaxis()
+            axis.xaxis.set_major_formatter(tick_formatters[metric])
+            axis.xaxis.set_tick_params(rotation=30)
+            axis.grid(True)
+            axis.spines['top'].set_visible(False)
+            axis.spines['right'].set_visible(False)
+            axis.set_axisbelow(True)
+
+            # force xmax to 100%
+            if metric in ['write_efficiency', 'read_efficiency']:
+                axis.set_xlim(right=1.0)
+
+            title = self.metric_formats[metric].name
+            axis.set_title(title, fontsize=12)
+
+            plt.tight_layout()
+            filename = '{}.{}'.format(metric, save_format)
+            fig.savefig(os.path.join(output_dir, filename), **savefig_kwargs)
+
+        if backend != 'Agg':
+            plt.show()
 
 # Run ----------------------------------------------------------------------------------------------
 
-def run_benchmark(args):
+def run_benchmark(cmd_args):
+    # run as separate process, because else we cannot capture all output from verilator
     benchmark_script = os.path.join(os.path.dirname(__file__), 'benchmark.py')
-    command = ['python3', benchmark_script, *args]
+    command = ['python3', benchmark_script, *cmd_args]
     proc = subprocess.run(command, capture_output=True, text=True, check=True)
     return proc.stdout
 
 
-def main():
+def run_benchmarks(configurations):
+    benchmarks = []
+    for name, config in configurations.items():
+        cmd_args = config.as_args()
+        print('{}: {}'.format(name, ' '.join(cmd_args)))
+        output = run_benchmark(cmd_args)
+
+        # return raw outputs, not BenchmarkResult so that we can store them in a file
+        benchmarks.append((config, output))
+
+        # exit if checker had any read error
+        result = BenchmarkResult(config, output)
+        if result.checker_errors != 0:
+            print('Error during benchmark "{}": checker_errors = {}'.format(
+                name, result.checker_errors), file=sys.stderr)
+            sys.exit(1)
+    return benchmarks
+
+
+def main(argv=None):
     parser = argparse.ArgumentParser(
-        description='Run LiteDRAM benchmarks and collect the results')
+        description='Run LiteDRAM benchmarks and collect the results.')
-    parser.add_argument('--yaml', required=True, help='Load benchmark configurations from YAML file')
+    parser.add_argument('--yaml',             required=True,       help='Load benchmark configurations from YAML file')
-    parser.add_argument('--names', nargs='*', help='Limit benchmarks to given names')
+    parser.add_argument('--names',            nargs='*',           help='Limit benchmarks to given names')
-    parser.add_argument('--regex', help='Limit benchmarks to names matching the regex')
+    parser.add_argument('--regex',                                 help='Limit benchmarks to names matching the regex')
-    parser.add_argument('--not-regex', help='Limit benchmarks to names not matching the regex')
+    parser.add_argument('--not-regex',                             help='Limit benchmarks to names not matching the regex')
-    args = parser.parse_args()
+    parser.add_argument('--plot',             action='store_true', help='Generate plots with results summary')
+    parser.add_argument('--plot-format',      default='png',       help='Specify plots file format (default=png)')
+    parser.add_argument('--plot-backend',     default='Agg',       help='Optionally specify matplotlib GUI backend')
+    parser.add_argument('--plot-transparent', action='store_true', help='Use transparent background when saving plots')
+    parser.add_argument('--plot-output-dir',  default='plots',     help='Specify where to save the plots')
+    parser.add_argument('--plot-theme',       default='default',   help='Use different matplotlib theme')
+    parser.add_argument('--output-cache',                          help='Cache benchmark outputs to given file if it exists, else load them from the file without running benchmarks. This allows to run the script multiple times to produce different outputs from the same run')
+    args = parser.parse_args(argv)
 
     # load and filter configurations
     configurations = BenchmarkConfiguration.load_yaml(args.yaml)
@@ -169,29 +289,37 @@ def main():
     for f in filters:
         configurations = dict(filter(f, configurations.items()))
 
-    # run the benchmarks
+    cache_exists = args.output_cache and os.path.isfile(args.output_cache)
-    results = []
-    for name, config in configurations.items():
-        args = config.as_args()
-        print('{}: {}'.format(name, ' '.join(args)))
 
-        result = BenchmarkResult(config, run_benchmark(args))
+    # load outputs from cache if it exsits
-        results.append(result)
+    if args.output_cache and cache_exists:
+        import pickle
+        with open(args.output_cache, 'rb') as f:
+            cached_benchmarks = pickle.load(f)
+        # take only those that match configurations
+        benchmarks = [(c, o) for c, o in cached_benchmarks if c in configurations.values()]
+    else:  # run all the benchmarks normally
+        benchmarks = run_benchmarks(configurations)
 
-        print("""\
+    # store outputs in cache
-  write_bandwidth  = {:6.3f} {}bps
+    if args.output_cache and not cache_exists:
-  read_bandwidth   = {:6.3f} {}bps
+        import pickle
-  write_efficiency = {:6.2f} %
+        with open(args.output_cache, 'wb') as f:
-  read_efficiency  = {:6.2f} %
+            pickle.dump(benchmarks, f, pickle.HIGHEST_PROTOCOL)
-        """.rstrip().format(
-            *human_readable(result.write_bandwidth()),
-            *human_readable(result.read_bandwidth()),
-            result.write_efficiency() * 100,
-            result.read_efficiency() * 100,
-        ))
 
+    # display the summary
+    results = [BenchmarkResult(config, output) for config, output in benchmarks]
     summary = ResultsSummary(results)
     summary.print()
+    if args.plot:
+        if not os.path.isdir(args.plot_output_dir):
+            os.makedirs(args.plot_output_dir)
+        summary.plot(args.plot_output_dir,
+                     backend=args.plot_backend,
+                     theme=args.plot_theme,
+                     save_format=args.plot_format,
+                     transparent=args.plot_transparent)
+
 
 if __name__ == "__main__":
     main()