Merge pull request #2372 from igchor/bench_cpu_count

[Benchmarks] support CPU counter measurements

Author: pbalcer

scripts/benchmarks/benches/base.py

@@ -57,9 +57,6 @@ def download(self, name, url, file, untar = False):
     def name(self):
         raise NotImplementedError()
 
-    def unit(self):
-        raise NotImplementedError()
-
     def lower_is_better(self):
         return True
 

scripts/benchmarks/benches/compute.py

@@ -19,7 +19,7 @@ def setup(self):
         if options.sycl is None:
             return
 
-        repo_path = git_clone(self.directory, "compute-benchmarks-repo", "https://github.com/intel/compute-benchmarks.git", "c80ddec9f0b4905bcbeb0f264f710093dc70340d")
+        repo_path = git_clone(self.directory, "compute-benchmarks-repo", "https://github.com/intel/compute-benchmarks.git", "df38bc342641d7e83fbb4fe764a23d21d734e07b")
         build_path = create_build_path(self.directory, 'compute-benchmarks-build')
 
         configure_command = [
@@ -77,6 +77,13 @@ def benchmarks(self) -> list[Benchmark]:
 
         return benches
 
+def parse_unit_type(compute_unit):
+    if "[count]" in compute_unit:
+        return "instr"
+    elif "[us]" in compute_unit:
+        return "μs"
+    return "unknown"
+
 class ComputeBenchmark(Benchmark):
     def __init__(self, bench, name, test):
         self.bench = bench
@@ -90,9 +97,6 @@ def bin_args(self) -> list[str]:
     def extra_env_vars(self) -> dict:
         return {}
 
-    def unit(self):
-        return "μs"
-
     def setup(self):
         self.benchmark_bin = os.path.join(self.bench.directory, 'compute-benchmarks-build', 'bin', self.bench_name)
 
@@ -108,22 +112,32 @@ def run(self, env_vars) -> list[Result]:
         env_vars.update(self.extra_env_vars())
 
         result = self.run_bench(command, env_vars)
-        (label, mean) = self.parse_output(result)
-        return [ Result(label=self.name(), value=mean, command=command, env=env_vars, stdout=result) ]
+        parsed_results = self.parse_output(result)
+        ret = []
+        for label, mean, unit in parsed_results:
+            extra_label = " CPU count" if parse_unit_type(unit) == "CPU count" else ""
+            ret.append(Result(label=self.name() + extra_label, value=mean, command=command, env=env_vars, stdout=result, unit=parse_unit_type(unit)))
+        return ret
 
     def parse_output(self, output):
         csv_file = io.StringIO(output)
         reader = csv.reader(csv_file)
         next(reader, None)
-        data_row = next(reader, None)
-        if data_row is None:
+        results = []
+        while True:
+            data_row = next(reader, None)
+            if data_row is None:
+                break
+            try:
+                label = data_row[0]
+                mean = float(data_row[1])
+                unit = data_row[7]
+                results.append((label, mean, unit))
+            except (ValueError, IndexError) as e:
+                raise ValueError(f"Error parsing output: {e}")
+        if len(results) == 0:
             raise ValueError("Benchmark output does not contain data.")
-        try:
-            label = data_row[0]
-            mean = float(data_row[1])
-            return (label, mean)
-        except (ValueError, IndexError) as e:
-            raise ValueError(f"Error parsing output: {e}")
+        return results
 
     def teardown(self):
         return
@@ -249,6 +263,7 @@ def bin_args(self) -> list[str]:
             f"--memoryPlacement={self.placement}",
             "--useEvents=0",
             "--contents=Zeros",
+            "--multiplier=1",
         ]
 
 class VectorSum(ComputeBenchmark):

scripts/benchmarks/benches/llamacpp.py

@@ -138,9 +138,6 @@ def __init__(self, bench):
         self.bench = bench
         super().__init__(bench.directory)
 
-    def unit(self):
-        return "token/s"
-
     def setup(self):
         self.benchmark_bin = os.path.join(self.bench.build_path, 'bin', 'llama-bench')
 
@@ -171,7 +168,7 @@ def run(self, env_vars) -> list[Result]:
         for r in parsed:
             (extra_label, mean) = r
             label = f"{self.name()} {extra_label}"
-            results.append(Result(label=label, value=mean, command=command, env=env_vars, stdout=result))
+            results.append(Result(label=label, value=mean, command=command, env=env_vars, stdout=result, unit="token/s"))
         return results
 
     def parse_output(self, output):

scripts/benchmarks/benches/result.py

@@ -17,8 +17,8 @@ class Result:
     env: str
     stdout: str
     passed: bool = True
-    # values should not be set by the benchmark
     unit: str = ""
+    # values should not be set by the benchmark
     name: str = ""
     lower_is_better: bool = True
     git_hash: str = ''

scripts/benchmarks/benches/syclbench.py

@@ -99,9 +99,6 @@ def bin_args(self) -> list[str]:
     def extra_env_vars(self) -> dict:
         return {}
 
-    def unit(self):
-        return "ms"
-
     def setup(self):
         self.benchmark_bin = os.path.join(self.directory, 'sycl-bench-build', self.bench_name)
 
@@ -134,7 +131,8 @@ def run(self, env_vars) -> list[Result]:
                             passed=(row[1]=="PASS"),
                             command=command,
                             env=env_vars,
-                            stdout=row))
+                            stdout=row,
+                            unit="ms"))
         self.done = True
         return res_list
 

scripts/benchmarks/benches/test.py

@@ -49,9 +49,6 @@ def __init__(self, name, value, diff):
     def name(self):
         return self.bname
 
-    def unit(self):
-        return "ms"
-
     def lower_is_better(self):
         return True
 
@@ -61,7 +58,7 @@ def setup(self):
     def run(self, env_vars) -> list[Result]:
         random_value = self.value + random.uniform(-1 * (self.diff), self.diff)
         return [
-            Result(label=self.name(), value=random_value, command="", env={"A": "B"}, stdout="no output")
+            Result(label=self.name(), value=random_value, command="", env={"A": "B"}, stdout="no output", unit="ms")
         ]
 
     def teardown(self):

scripts/benchmarks/benches/velocity.py

@@ -39,11 +39,12 @@ def benchmarks(self) -> list[Benchmark]:
         ]
 
 class VelocityBase(Benchmark):
-    def __init__(self, name: str, bin_name: str, vb: VelocityBench):
+    def __init__(self, name: str, bin_name: str, vb: VelocityBench, unit: str):
         super().__init__(vb.directory)
         self.vb = vb
         self.bench_name = name
         self.bin_name = bin_name
+        self.unit = unit
         self.code_path = os.path.join(self.vb.repo_path, self.bench_name, 'SYCL')
 
     def download_deps(self):
@@ -83,21 +84,18 @@ def run(self, env_vars) -> list[Result]:
 
         result = self.run_bench(command, env_vars)
 
-        return [ Result(label=self.name(), value=self.parse_output(result), command=command, env=env_vars, stdout=result) ]
+        return [ Result(label=self.name(), value=self.parse_output(result), command=command, env=env_vars, stdout=result, unit=self.unit) ]
 
     def teardown(self):
         return
 
 class Hashtable(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        super().__init__("hashtable", "hashtable_sycl", vb)
+        super().__init__("hashtable", "hashtable_sycl", vb, "M keys/sec")
 
     def name(self):
         return "Velocity-Bench Hashtable"
 
-    def unit(self):
-        return "M keys/sec"
-
     def bin_args(self) -> list[str]:
         return ["--no-verify"]
 
@@ -114,15 +112,12 @@ def parse_output(self, stdout: str) -> float:
 
 class Bitcracker(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        super().__init__("bitcracker", "bitcracker", vb)
+        super().__init__("bitcracker", "bitcracker", vb, "s")
         self.data_path = os.path.join(vb.repo_path, "bitcracker", "hash_pass")
 
     def name(self):
         return "Velocity-Bench Bitcracker"
 
-    def unit(self):
-        return "s"
-
     def bin_args(self) -> list[str]:
         return ["-f", f"{self.data_path}/img_win8_user_hash.txt",
                 "-d", f"{self.data_path}/user_passwords_60000.txt",
@@ -137,7 +132,7 @@ def parse_output(self, stdout: str) -> float:
 
 class SobelFilter(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        super().__init__("sobel_filter", "sobel_filter", vb)
+        super().__init__("sobel_filter", "sobel_filter", vb, "ms")
 
     def download_deps(self):
         self.download("sobel_filter", "https://github.com/oneapi-src/Velocity-Bench/raw/main/sobel_filter/res/sobel_filter_data.tgz?download=", "sobel_filter_data.tgz", untar=True)
@@ -146,9 +141,6 @@ def download_deps(self):
     def name(self):
         return "Velocity-Bench Sobel Filter"
 
-    def unit(self):
-        return "ms"
-
     def bin_args(self) -> list[str]:
         return ["-i", f"{self.data_path}/sobel_filter_data/silverfalls_32Kx32K.png",
                 "-n", "5"]
@@ -166,7 +158,7 @@ def parse_output(self, stdout: str) -> float:
 
 class QuickSilver(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        super().__init__("QuickSilver", "qs", vb)
+        super().__init__("QuickSilver", "qs", vb, "MMS/CTT")
         self.data_path = os.path.join(vb.repo_path, "QuickSilver", "Examples", "AllScattering")
 
     def run(self, env_vars) -> list[Result]:
@@ -179,9 +171,6 @@ def run(self, env_vars) -> list[Result]:
     def name(self):
         return "Velocity-Bench QuickSilver"
 
-    def unit(self):
-        return "MMS/CTT"
-
     def lower_is_better(self):
         return False
 
@@ -200,17 +189,14 @@ def parse_output(self, stdout: str) -> float:
 
 class Easywave(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        super().__init__("easywave", "easyWave_sycl", vb)
+        super().__init__("easywave", "easyWave_sycl", vb, "ms")
 
     def download_deps(self):
         self.download("easywave", "https://git.gfz-potsdam.de/id2/geoperil/easyWave/-/raw/master/data/examples.tar.gz", "examples.tar.gz", untar=True)
 
     def name(self):
         return "Velocity-Bench Easywave"
 
-    def unit(self):
-        return "ms"
-
     def bin_args(self) -> list[str]:
         return ["-grid", f"{self.data_path}/examples/e2Asean.grd",
                 "-source", f"{self.data_path}/examples/BengkuluSept2007.flt",
@@ -245,7 +231,7 @@ def parse_output(self, stdout: str) -> float:
 
 class CudaSift(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        super().__init__("cudaSift", "cudaSift", vb)
+        super().__init__("cudaSift", "cudaSift", vb, "ms")
 
     def download_deps(self):
         images = os.path.join(self.vb.repo_path, self.bench_name, 'inputData')
@@ -256,9 +242,6 @@ def download_deps(self):
     def name(self):
         return "Velocity-Bench CudaSift"
 
-    def unit(self):
-        return "ms"
-
     def parse_output(self, stdout: str) -> float:
         match = re.search(r'Avg workload time = (\d+\.\d+) ms', stdout)
         if match:

scripts/benchmarks/main.py

@@ -73,7 +73,7 @@ def main(directory, additional_env_vars, save_name, compare_names, filter):
                 if bench_results is not None:
                     for bench_result in bench_results:
                         if bench_result.passed:
-                            print(f"complete ({bench_result.label}: {bench_result.value:.3f} {benchmark.unit()}).")
+                            print(f"complete ({bench_result.label}: {bench_result.value:.3f} {bench_result.unit}).")
                         else:
                             print(f"complete ({bench_result.label}: verification FAILED)")
                         iteration_results.append(bench_result)
@@ -91,7 +91,6 @@ def main(directory, additional_env_vars, save_name, compare_names, filter):
                     median_index = len(label_results) // 2
                     median_result = label_results[median_index]
 
-                    median_result.unit = benchmark.unit()
                     median_result.name = label
                     median_result.lower_is_better = benchmark.lower_is_better()