[UR][Benchmarks] Support for Unitrace for all benchmarks

devops/scripts/benchmarks/benches/base.py

@@ -3,7 +3,6 @@
 # See LICENSE.TXT
 # SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 
-from dataclasses import dataclass
 import os
 import shutil
 import subprocess
@@ -12,6 +11,8 @@
 from options import options
 from utils.utils import download, run
 from abc import ABC, abstractmethod
+from utils.unitrace import get_unitrace
+from utils.logger import log
 
 benchmark_tags = [
     BenchmarkTag("SYCL", "Benchmark uses SYCL runtime"),
@@ -61,6 +62,12 @@ 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 Unitrace.
+        By default, it returns True, but can be overridden to disable tracing for a benchmark.
+        """
+        return True
+
     @abstractmethod
     def setup(self):
         pass
@@ -70,11 +77,12 @@ def teardown(self):
         pass
 
     @abstractmethod
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_unitrace: bool = False) -> list[Result]:
         """Execute the benchmark with the given environment variables.
 
         Args:
             env_vars: Environment variables to use when running the benchmark.
+            run_unitrace: Whether to run benchmark under Unitrace.
 
         Returns:
             A list of Result objects with the benchmark results.
@@ -97,7 +105,14 @@ 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_unitrace=False,
+        extra_unitrace_opt=None,
     ):
         env_vars = env_vars.copy()
         if options.ur is not None:
@@ -110,13 +125,30 @@ 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,
-        )
+        if self.traceable() and run_unitrace:
+            if extra_unitrace_opt is None:
+                extra_unitrace_opt = []
+            unitrace_output, command = get_unitrace().setup(
+                self.name(), command, extra_unitrace_opt
+            )
+            log.debug(f"Unitrace output: {unitrace_output}")
+            log.debug(f"Unitrace 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_unitrace:
+                get_unitrace().cleanup(options.benchmark_cwd, unitrace_output)
+            raise
+
+        if self.traceable() and run_unitrace:
+            get_unitrace().handle_output(unitrace_output)
 
         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_unitrace: bool = False) -> list[Result]:
         command = [
             str(self.bench_bin),
             *self.bench_args.split(),
@@ -151,6 +151,8 @@ def run(self, env_vars):
             add_sycl=True,
             ld_library=ld_library,
             use_stdout=True,
+            run_unitrace=run_unitrace,
+            extra_unitrace_opt=["--chrome-dnn-logging"],
         )
         result_value = self._extract_time(output)
 

devops/scripts/benchmarks/benches/benchdnn_list.py

@@ -62,6 +62,7 @@
         "graph",
         "sdpa-plain-f16",
         "--reset --dt=f16 --case=complex_fusion/mha/sdpa-plain-implicit-causal-mask-fp32-bs1.json",
+        False,  # Do not run SYCL graph for this benchmark
     ],
     [
         "graph",

devops/scripts/benchmarks/benches/compute.py

@@ -303,6 +303,10 @@ def enabled(self) -> bool:
         # Check if the specific runtime is enabled (or no specific runtime required)
         return self.runtime is None or self.runtime in self.enabled_runtimes()
 
+    def name(self):
+        """Returns the name of the benchmark, can be overridden."""
+        return self.bench_name
+
     def bin_args(self) -> list[str]:
         return []
 
@@ -320,7 +324,7 @@ def explicit_group(self):
     def description(self) -> str:
         return ""
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_unitrace: bool = False) -> list[Result]:
         command = [
             f"{self.benchmark_bin}",
             f"--test={self.test}",
@@ -331,7 +335,11 @@ 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_unitrace=run_unitrace,
+        )
         parsed_results = self.parse_output(result)
         ret = []
         for label, median, stddev, unit in parsed_results:

devops/scripts/benchmarks/benches/gromacs.py

@@ -13,6 +13,7 @@
 from utils.result import Result
 from utils.oneapi import get_oneapi
 from utils.logger import log
+from utils.unitrace import get_unitrace
 
 
 class GromacsBench(Suite):
@@ -64,24 +65,29 @@ def setup(self) -> None:
 
         self.oneapi = get_oneapi()
 
+        cmd_list = [
+            "cmake",
+            f"-S {self.gromacs_src}",
+            f"-B {self.gromacs_build_path}",
+            "-DCMAKE_BUILD_TYPE=Release",
+            "-DCMAKE_CXX_COMPILER=clang++",
+            "-DCMAKE_C_COMPILER=clang",
+            "-DGMX_GPU=SYCL",
+            "-DGMX_SYCL_ENABLE_GRAPHS=ON",
+            "-DGMX_SYCL_ENABLE_EXPERIMENTAL_SUBMIT_API=ON",
+            "-DGMX_FFT_LIBRARY=MKL",
+            "-DGMX_GPU_FFT_LIBRARY=MKL",
+            f"-DMKLROOT={self.oneapi.mkl_dir()}",
+            "-DGMX_GPU_NB_CLUSTER_SIZE=8",
+            "-DGMX_GPU_NB_NUM_CLUSTER_PER_CELL_X=1",
+            "-DGMX_OPENMP=OFF",
+        ]
+
+        if options.unitrace:
+            cmd_list.append("-DGMX_USE_ITT=ON")
+
         run(
-            [
-                "cmake",
-                f"-S {str(self.directory)}/gromacs-repo",
-                f"-B {self.gromacs_build_path}",
-                f"-DCMAKE_BUILD_TYPE=Release",
-                f"-DCMAKE_CXX_COMPILER=clang++",
-                f"-DCMAKE_C_COMPILER=clang",
-                f"-DGMX_GPU=SYCL",
-                f"-DGMX_SYCL_ENABLE_GRAPHS=ON",
-                f"-DGMX_SYCL_ENABLE_EXPERIMENTAL_SUBMIT_API=ON",
-                f"-DGMX_FFT_LIBRARY=MKL",
-                f"-DGMX_GPU_FFT_LIBRARY=MKL",
-                f"-DMKLROOT={self.oneapi.mkl_dir()}",
-                f"-DGMX_GPU_NB_CLUSTER_SIZE=8",
-                f"-DGMX_GPU_NB_NUM_CLUSTER_PER_CELL_X=1",
-                f"-DGMX_OPENMP=OFF",
-            ],
+            cmd_list,
             add_sycl=True,
         )
         run(
@@ -163,7 +169,7 @@ def setup(self):
             ld_library=self.suite.oneapi.ld_libraries(),
         )
 
-    def run(self, env_vars):
+    def run(self, env_vars, run_unitrace: bool = False) -> list[Result]:
         model_dir = self.grappa_dir / self.model
 
         env_vars.update({"SYCL_CACHE_PERSISTENT": "1"})
@@ -202,6 +208,7 @@ def run(self, env_vars):
             add_sycl=True,
             use_stdout=False,
             ld_library=self.suite.oneapi.ld_libraries(),
+            run_unitrace=run_unitrace,
         )
 
         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_unitrace: bool = False) -> list[Result]:
         command = [
             f"{self.benchmark_bin}",
             "--output",
@@ -141,7 +141,10 @@ 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_unitrace=run_unitrace,
         )
         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_unitrace: bool = False) -> list[Result]:
         self.outputfile = os.path.join(self.bench.directory, self.test + ".csv")
 
         command = [
@@ -151,7 +151,11 @@ 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_unitrace=run_unitrace,
+        )
 
         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_unitrace: 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_unitrace: bool = False) -> list[Result]:
         command = [f"{self.benchmark_bin}"]
 
         all_names = self.get_names_of_benchmarks_to_be_run(command, env_vars)
@@ -151,7 +151,11 @@ 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_unitrace=run_unitrace,
             )
 
             parsed = self.parse_output(result)

devops/scripts/benchmarks/benches/velocity.py

@@ -130,15 +130,20 @@ def description(self) -> str:
     def get_tags(self):
         return ["SYCL", "application"]
 
-    def run(self, env_vars) -> list[Result]:
+    def run(self, env_vars, run_unitrace: 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_unitrace=run_unitrace,
+        )
 
         return [
             Result(
@@ -282,7 +287,7 @@ 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_unitrace: 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

devops/scripts/benchmarks/history.py

@@ -15,11 +15,11 @@
 from utils.validate import Validate
 from utils.logger import log
 from utils.detect_versions import DetectVersion
+from utils.unitrace import get_unitrace
 
 
 class BenchmarkHistory:
     runs = []
-    TIMESTAMP_FORMAT = "%Y%m%d_%H%M%S"
 
     def __init__(self, dir):
         self.dir = dir
@@ -78,7 +78,9 @@ def is_file_too_old(file_path: Path) -> bool:
                 if not timestamp_str:
                     return False
 
-                file_timestamp = datetime.strptime(timestamp_str, self.TIMESTAMP_FORMAT)
+                file_timestamp = datetime.strptime(
+                    timestamp_str, options.TIMESTAMP_FORMAT
+                )
                 # Add timezone info for proper comparison
                 file_timestamp = file_timestamp.replace(tzinfo=timezone.utc)
                 return file_timestamp < cutoff_date
@@ -190,23 +192,28 @@ def git_info_from_path(path: Path) -> (str, str):
             compute_runtime=compute_runtime,
         )
 
-    def save(self, save_name, results: list[Result], to_file=True):
+    def save(self, save_name, results: list[Result]):
         benchmark_data = self.create_run(save_name, results)
         self.runs.append(benchmark_data)
 
-        if not to_file:
+        if options.save_name is None:
             return
 
-        serialized = benchmark_data.to_json()
+        serialized = benchmark_data.to_json()  # type: ignore
         results_dir = Path(os.path.join(self.dir, "results"))
         os.makedirs(results_dir, exist_ok=True)
 
-        # Use formatted timestamp for the filename
-        timestamp = (
-            datetime.now(tz=timezone.utc).strftime(self.TIMESTAMP_FORMAT)
-            if options.timestamp_override is None
-            else options.timestamp_override
-        )
+        if options.unitrace:
+            timestamp = get_unitrace().timestamp  # type: ignore
+        elif options.timestamp_override is not None:
+            timestamp = options.timestamp_override
+        else:
+            timestamp = (
+                datetime.now(tz=timezone.utc).strftime(self.TIMESTAMP_FORMAT)
+                if options.timestamp_override is None
+                else options.timestamp_override
+            )
+
         file_path = Path(os.path.join(results_dir, f"{save_name}_{timestamp}.json"))
         with file_path.open("w") as file:
             json.dump(serialized, file, indent=4)