Add flamegraphs to benchmarks

Signed-off-by: Mateusz P. Nowak <[email protected]>

Author: mateuszpn

devops/scripts/benchmarks/benches/base.py

@@ -12,6 +12,8 @@
 from options import options
 from utils.utils import download, run
 from abc import ABC, abstractmethod
+from utils.flamegraph import get_flamegraph
+from utils.logger import log
 
 benchmark_tags = [
     BenchmarkTag("SYCL", "Benchmark uses SYCL runtime"),
@@ -61,6 +63,12 @@ def enabled(self) -> bool:
         By default, it returns True, but can be overridden to disable a benchmark."""
         return True
 
+    def traceable(self) -> bool:
+        """Returns whether this benchmark should be traced by FlameGraph.
+        By default, it returns True, but can be overridden to disable tracing for a benchmark.
+        """
+        return True
+
     @abstractmethod
     def setup(self):
         pass
@@ -70,11 +78,12 @@ def teardown(self):
         pass
 
     @abstractmethod
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         """Execute the benchmark with the given environment variables.
 
         Args:
             env_vars: Environment variables to use when running the benchmark.
+            run_flamegraph: Whether to run benchmark under FlameGraph.
 
         Returns:
             A list of Result objects with the benchmark results.
@@ -97,7 +106,14 @@ def get_adapter_full_path():
         ), f"could not find adapter file {adapter_path} (and in similar lib paths)"
 
     def run_bench(
-        self, command, env_vars, ld_library=[], add_sycl=True, use_stdout=True
+        self,
+        command,
+        env_vars,
+        ld_library=[],
+        add_sycl=True,
+        use_stdout=True,
+        run_flamegraph=False,
+        extra_perf_opt=None,
     ):
         env_vars = env_vars.copy()
         if options.ur is not None:
@@ -110,13 +126,32 @@ def run_bench(
         ld_libraries = options.extra_ld_libraries.copy()
         ld_libraries.extend(ld_library)
 
-        result = run(
-            command=command,
-            env_vars=env_vars,
-            add_sycl=add_sycl,
-            cwd=options.benchmark_cwd,
-            ld_library=ld_libraries,
-        )
+        perf_data_file = None
+        if self.traceable() and run_flamegraph:
+            if extra_perf_opt is None:
+                extra_perf_opt = []
+            perf_data_file, command = get_flamegraph().setup(
+                self.name(), command, extra_perf_opt
+            )
+            log.debug(f"FlameGraph perf data: {perf_data_file}")
+            log.debug(f"FlameGraph command: {' '.join(command)}")
+
+        try:
+            result = run(
+                command=command,
+                env_vars=env_vars,
+                add_sycl=add_sycl,
+                cwd=options.benchmark_cwd,
+                ld_library=ld_libraries,
+            )
+        except subprocess.CalledProcessError:
+            if run_flamegraph and perf_data_file:
+                get_flamegraph().cleanup(options.benchmark_cwd, perf_data_file)
+            raise
+
+        if self.traceable() and run_flamegraph and perf_data_file:
+            svg_file = get_flamegraph().handle_output(self.name(), perf_data_file)
+            log.info(f"FlameGraph generated: {svg_file}")
 
         if use_stdout:
             return result.stdout.decode()

devops/scripts/benchmarks/benches/benchdnn.py

@@ -132,7 +132,7 @@ def setup(self):
         if not self.bench_bin.exists():
             raise FileNotFoundError(f"Benchmark binary not found: {self.bench_bin}")
 
-    def run(self, env_vars):
+    def run(self, env_vars, run_flamegraph: bool = False):
         command = [
             str(self.bench_bin),
             *self.bench_args.split(),
@@ -151,6 +151,7 @@ def run(self, env_vars):
             add_sycl=True,
             ld_library=ld_library,
             use_stdout=True,
+            run_flamegraph=run_flamegraph,
         )
         result_value = self._extract_time(output)
 

devops/scripts/benchmarks/benches/benchdnn_list.py

@@ -14,29 +14,29 @@
 
 # the final choice of benchmarks to run, used in CI and other environments
 benches_final_set = [
-    [
-        "sum",
-        "f16-1",
-        "--sdt=f16:f16:f16 --stag=abx:abx:abx --scales=1.25:3:0.5 16x2x6x4x3",
-        False,  # Do not run graph for this benchmark
-    ],
-    [
-        "sum",
-        "f16-2",
-        "--reset --ddt=f16 \
-            --sdt=f16:f16:f16:f16:f16:f16:f16:f16:f16:f16 \
-            --stag=abx:aBx16b:ABx16a16b:ABcd16b16a:BAcd16a16b:BAcd16b16a:aBCd16b16c:aBCd16c16b:aCBd16b16c:aCBd16c16b \
-            --dtag=abx,aBx16b,ABx16a16b,ABcd16b16a,BAcd16a16b,BAcd16b16a,aBCd16b16c,aBCd16c16b,aCBd16b16c,aCBd16c16b \
-            --scales=1.25:3:0.5:2:0.5:2:0.5:2:0.5:2 \
-            16x32x48x5",
-        False,  # Do not run graph for this benchmark
-    ],
-    [
-        "sum",
-        "f32-1",
-        "--sdt=bf16:bf16:bf16 --stag=abx:abx:abx --scales=0.5:2:0.5    16x2x6x4x3",
-        False,  # Do not run graph for this benchmark
-    ],
+    # [
+    #     "sum",
+    #     "f16-1",
+    #     "--sdt=f16:f16:f16 --stag=abx:abx:abx --scales=1.25:3:0.5 16x2x6x4x3",
+    #     False,  # Do not run graph for this benchmark
+    # ],
+    # [
+    #     "sum",
+    #     "f16-2",
+    #     "--reset --ddt=f16 \
+    #         --sdt=f16:f16:f16:f16:f16:f16:f16:f16:f16:f16 \
+    #         --stag=abx:aBx16b:ABx16a16b:ABcd16b16a:BAcd16a16b:BAcd16b16a:aBCd16b16c:aBCd16c16b:aCBd16b16c:aCBd16c16b \
+    #         --dtag=abx,aBx16b,ABx16a16b,ABcd16b16a,BAcd16a16b,BAcd16b16a,aBCd16b16c,aBCd16c16b,aCBd16b16c,aCBd16c16b \
+    #         --scales=1.25:3:0.5:2:0.5:2:0.5:2:0.5:2 \
+    #         16x32x48x5",
+    #     False,  # Do not run graph for this benchmark
+    # ],
+    # [
+    #     "sum",
+    #     "f32-1",
+    #     "--sdt=bf16:bf16:bf16 --stag=abx:abx:abx --scales=0.5:2:0.5    16x2x6x4x3",
+    #     False,  # Do not run graph for this benchmark
+    # ],
     [
         "sum",
         "f32-2",
@@ -53,21 +53,21 @@
         "padding-1",
         "--ddt=f32 --sdt=f32:f32 --stag=aBcd16b --dtag=aBcd16b 1x8x64x64 1x8x640x1024 1x24x640x1024",
     ],
-    [
-        "sum",
-        "padding-2",
-        "--sdt=bf16:bf16 --ddt=bf16 --stag=AB48a16b:AB48a16b --dtag=AB48a16b 512x1024",
-    ],
-    [
-        "graph",
-        "sdpa-plain-f16",
-        "--reset --dt=f16 --case=complex_fusion/mha/sdpa-plain-implicit-causal-mask-fp32-bs1.json",
-    ],
-    [
-        "graph",
-        "sdpa-plain-f32",
-        "--reset --dt=f32 --case=complex_fusion/mha/sdpa-plain-implicit-causal-mask-fp32-bs1.json",
-    ],
+    # [
+    #     "sum",
+    #     "padding-2",
+    #     "--sdt=bf16:bf16 --ddt=bf16 --stag=AB48a16b:AB48a16b --dtag=AB48a16b 512x1024",
+    # ],
+    # [
+    #     "graph",
+    #     "sdpa-plain-f16",
+    #     "--reset --dt=f16 --case=complex_fusion/mha/sdpa-plain-implicit-causal-mask-fp32-bs1.json",
+    # ],
+    # [
+    #     "graph",
+    #     "sdpa-plain-f32",
+    #     "--reset --dt=f32 --case=complex_fusion/mha/sdpa-plain-implicit-causal-mask-fp32-bs1.json",
+    # ],
 ]
 
 # the complete set of benchmarks aimed at gpu operations, normally too long to run in CI

devops/scripts/benchmarks/benches/compute.py

@@ -320,7 +320,7 @@ def explicit_group(self):
     def description(self) -> str:
         return ""
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         command = [
             f"{self.benchmark_bin}",
             f"--test={self.test}",
@@ -331,7 +331,7 @@ def run(self, env_vars) -> list[Result]:
         command += self.bin_args()
         env_vars.update(self.extra_env_vars())
 
-        result = self.run_bench(command, env_vars)
+        result = self.run_bench(command, env_vars, run_flamegraph=run_flamegraph)
         parsed_results = self.parse_output(result)
         ret = []
         for label, median, stddev, unit in parsed_results:

devops/scripts/benchmarks/benches/gromacs.py

@@ -163,7 +163,7 @@ def setup(self):
             ld_library=self.suite.oneapi.ld_libraries(),
         )
 
-    def run(self, env_vars):
+    def run(self, env_vars, run_flamegraph: bool = False):
         model_dir = self.grappa_dir / self.model
 
         env_vars.update({"SYCL_CACHE_PERSISTENT": "1"})
@@ -202,6 +202,7 @@ def run(self, env_vars):
             add_sycl=True,
             use_stdout=False,
             ld_library=self.suite.oneapi.ld_libraries(),
+            run_flamegraph=run_flamegraph,
         )
 
         if not self._validate_correctness(options.benchmark_cwd + "/md.log"):

devops/scripts/benchmarks/benches/llamacpp.py

@@ -115,7 +115,7 @@ def get_tags(self):
     def lower_is_better(self):
         return False
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         command = [
             f"{self.benchmark_bin}",
             "--output",
@@ -141,7 +141,7 @@ def run(self, env_vars) -> list[Result]:
         ]
 
         result = self.run_bench(
-            command, env_vars, ld_library=self.bench.oneapi.ld_libraries()
+            command, env_vars, ld_library=self.bench.oneapi.ld_libraries(), run_flamegraph=run_flamegraph
         )
         parsed = self.parse_output(result)
         results = []

devops/scripts/benchmarks/benches/syclbench.py

@@ -137,7 +137,7 @@ def setup(self):
             self.directory, "sycl-bench-build", self.bench_name
         )
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         self.outputfile = os.path.join(self.bench.directory, self.test + ".csv")
 
         command = [
@@ -151,7 +151,7 @@ def run(self, env_vars) -> list[Result]:
         env_vars.update(self.extra_env_vars())
 
         # no output to stdout, all in outputfile
-        self.run_bench(command, env_vars)
+        self.run_bench(command, env_vars, run_flamegraph=run_flamegraph)
 
         with open(self.outputfile, "r") as f:
             reader = csv.reader(f)

devops/scripts/benchmarks/benches/test.py

@@ -88,7 +88,7 @@ def notes(self) -> str:
     def unstable(self) -> str:
         return self.unstable_text
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         random_value = self.value + random.uniform(-1 * (self.diff), self.diff)
         return [
             Result(

devops/scripts/benchmarks/benches/umf.py

@@ -137,7 +137,7 @@ def get_names_of_benchmarks_to_be_run(self, command, env_vars):
 
         return all_names
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         command = [f"{self.benchmark_bin}"]
 
         all_names = self.get_names_of_benchmarks_to_be_run(command, env_vars)
@@ -151,7 +151,7 @@ def run(self, env_vars) -> list[Result]:
             specific_benchmark = command + ["--benchmark_filter=^" + name + "$"]
 
             result = self.run_bench(
-                specific_benchmark, env_vars, add_sycl=False, ld_library=[self.umf_lib]
+                specific_benchmark, env_vars, add_sycl=False, ld_library=[self.umf_lib], run_flamegraph=run_flamegraph
             )
 
             parsed = self.parse_output(result)

devops/scripts/benchmarks/benches/velocity.py

@@ -130,15 +130,15 @@ def description(self) -> str:
     def get_tags(self):
         return ["SYCL", "application"]
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         env_vars.update(self.extra_env_vars())
 
         command = [
             f"{self.benchmark_bin}",
         ]
         command += self.bin_args()
 
-        result = self.run_bench(command, env_vars, ld_library=self.ld_libraries())
+        result = self.run_bench(command, env_vars, ld_library=self.ld_libraries(), run_flamegraph=run_flamegraph)
 
         return [
             Result(
@@ -282,7 +282,7 @@ class QuickSilver(VelocityBase):
     def __init__(self, vb: VelocityBench):
         super().__init__("QuickSilver", "qs", vb, "MMS/CTT")
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_flamegraph: bool = False) -> list[Result]:
         # TODO: fix the crash in QuickSilver when UR_L0_USE_IMMEDIATE_COMMANDLISTS=0
         if (
             "UR_L0_USE_IMMEDIATE_COMMANDLISTS" in env_vars