Merge pull request #2577 from oneapi-src/bench-build-umd

add building compute-runtime UMD in benchmarks jobs

Author: igchor

.github/workflows/benchmarks-reusable.yml

@@ -34,6 +34,10 @@ on:
         required: false
         type: boolean
         default: false
+      compute_runtime_commit:
+        required: false
+        type: string
+        default: ''
 
 permissions:
   contents: read
@@ -200,6 +204,7 @@ jobs:
         --ur ${{ github.workspace }}/ur_install
         --umf ${{ github.workspace }}/umf_build
         --adapter ${{ matrix.adapter.str_name }}
+        --compute-runtime ${{ inputs.compute_runtime_commit }}
         ${{ inputs.upload_report && '--output-html' || '' }}
         ${{ inputs.bench_script_params }}
 

.github/workflows/benchmarks.yml

@@ -43,6 +43,11 @@ on:
         type: string
         required: false
         default: ''
+      compute_runtime_commit:
+        description: 'Compute Runtime commit'
+        type: string
+        required: false
+        default: ''
       upload_report:
         description: 'Upload HTML report'
         type: boolean
@@ -65,4 +70,5 @@ jobs:
       sycl_config_params: ${{ inputs.sycl_config_params }}
       sycl_repo: ${{ inputs.sycl_repo }}
       sycl_commit: ${{ inputs.sycl_commit }}
+      compute_runtime_commit: ${{ inputs.compute_runtime_commit }}
       upload_report: ${{ inputs.upload_report }}

scripts/benchmarks/benches/base.py

@@ -7,7 +7,7 @@
 import shutil
 from pathlib import Path
 from .result import Result
-from .options import options
+from options import options
 from utils.utils import download, run
 import urllib.request
 import tarfile
@@ -28,17 +28,22 @@ 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):
-        env_vars_with_forced_adapter = env_vars.copy()
+        env_vars = env_vars.copy()
         if options.ur is not None:
-            env_vars_with_forced_adapter.update(
+            env_vars.update(
                 {'UR_ADAPTERS_FORCE_LOAD': Benchmark.get_adapter_full_path()})
 
+        env_vars.update(options.extra_env_vars)
+
+        ld_libraries = options.extra_ld_libraries.copy()
+        ld_libraries.extend(ld_library)
+
         return run(
             command=command,
-            env_vars=env_vars_with_forced_adapter,
+            env_vars=env_vars,
             add_sycl=add_sycl,
             cwd=options.benchmark_cwd,
-            ld_library=ld_library
+            ld_library=ld_libraries
         ).stdout.decode()
 
     def create_data_path(self, name, skip_data_dir = False):

scripts/benchmarks/benches/compute.py

@@ -9,7 +9,7 @@
 from utils.utils import run, git_clone, create_build_path
 from .base import Benchmark, Suite
 from .result import Result
-from .options import options
+from options import options
 
 class ComputeBench(Suite):
     def __init__(self, directory):
@@ -22,7 +22,7 @@ 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", "df38bc342641d7e83fbb4fe764a23d21d734e07b")
+        repo_path = git_clone(self.directory, "compute-benchmarks-repo", "https://github.com/intel/compute-benchmarks.git", "d13e5b4d8dd3d28926a74ab7f67f78c10f708a01")
         build_path = create_build_path(self.directory, 'compute-benchmarks-build')
 
         configure_command = [

scripts/benchmarks/benches/llamacpp.py

@@ -10,7 +10,7 @@
 from .base import Benchmark, Suite
 from .result import Result
 from utils.utils import run, create_build_path
-from .options import options
+from options import options
 from .oneapi import get_oneapi
 import os
 

scripts/benchmarks/benches/oneapi.py

@@ -5,7 +5,7 @@
 
 from pathlib import Path
 from utils.utils import download, run
-from .options import options
+from options import options
 import os
 
 class OneAPI:

scripts/benchmarks/benches/syclbench.py

@@ -9,7 +9,7 @@
 from utils.utils import run, git_clone, create_build_path
 from .base import Benchmark, Suite
 from .result import Result
-from .options import options
+from options import options
 
 class SyclBench(Suite):
     def __init__(self, directory):

scripts/benchmarks/benches/test.py

@@ -8,7 +8,7 @@
 from .base import Benchmark, Suite
 from .result import Result
 from utils.utils import run, create_build_path
-from .options import options
+from options import options
 import os
 
 class TestSuite(Suite):

scripts/benchmarks/benches/umf.py

@@ -8,7 +8,7 @@
 from .base import Benchmark, Suite
 from .result import Result
 from utils.utils import run, create_build_path
-from .options import options
+from options import options
 from .oneapi import get_oneapi
 import os
 import csv

scripts/benchmarks/benches/velocity.py

@@ -9,7 +9,7 @@
 from .base import Benchmark, Suite
 from .result import Result
 from utils.utils import run, create_build_path
-from .options import options
+from options import options
 from .oneapi import get_oneapi
 import shutil
 
@@ -54,7 +54,6 @@ def __init__(self, name: str, bin_name: str, vb: VelocityBench, unit: str):
         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):
         return
@@ -66,6 +65,7 @@ def ld_libraries(self) -> list[str]:
         return []
 
     def setup(self):
+        self.code_path = os.path.join(self.vb.repo_path, self.bench_name, 'SYCL')
         self.download_deps()
         self.benchmark_bin = os.path.join(self.directory, self.bench_name, self.bin_name)
 
@@ -130,12 +130,13 @@ def parse_output(self, stdout: str) -> float:
 class Bitcracker(VelocityBase):
     def __init__(self, vb: VelocityBench):
         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 bin_args(self) -> list[str]:
+        self.data_path = os.path.join(self.vb.repo_path, "bitcracker", "hash_pass")
+
         return ["-f", f"{self.data_path}/img_win8_user_hash.txt",
                 "-d", f"{self.data_path}/user_passwords_60000.txt",
                 "-b", "60000"]
@@ -175,7 +176,6 @@ def parse_output(self, stdout: str) -> float:
 class QuickSilver(VelocityBase):
     def __init__(self, vb: VelocityBench):
         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]:
         # TODO: fix the crash in QuickSilver when UR_L0_USE_IMMEDIATE_COMMANDLISTS=0
@@ -191,6 +191,8 @@ def lower_is_better(self):
         return False
 
     def bin_args(self) -> list[str]:
+        self.data_path = os.path.join(self.vb.repo_path, "QuickSilver", "Examples", "AllScattering")
+
         return ["-i", f"{self.data_path}/scatteringOnly.inp"]
 
     def extra_env_vars(self) -> dict:
@@ -266,20 +268,20 @@ def parse_output(self, stdout: str) -> float:
 
 class DLCifar(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        self.oneapi = get_oneapi()
         super().__init__("dl-cifar", "dl-cifar_sycl", vb, "s")
 
     def ld_libraries(self):
-        return self.oneapi.ld_libraries()
+        return get_oneapi().ld_libraries()
 
     def download_deps(self):
         # TODO: dl-cifar hardcodes the path to this dataset as "../../datasets/cifar-10-binary"...
         self.download("datasets", "https://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz", "cifar-10-binary.tar.gz", untar=True, skip_data_dir=True)
         return
 
     def extra_cmake_args(self):
+        oneapi = get_oneapi()
         return [
-            f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{self.oneapi.dnn_include()} -I{self.oneapi.mkl_include()} -L{self.oneapi.dnn_lib()} -L{self.oneapi.mkl_lib()}"
+            f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}"
         ]
 
     def name(self):
@@ -294,11 +296,10 @@ def parse_output(self, stdout: str) -> float:
 
 class DLMnist(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        self.oneapi = get_oneapi()
         super().__init__("dl-mnist", "dl-mnist-sycl", vb, "s")
 
     def ld_libraries(self):
-        return self.oneapi.ld_libraries()
+        return get_oneapi().ld_libraries()
 
     def download_deps(self):
         # TODO: dl-mnist hardcodes the path to this dataset as "../../datasets/"...
@@ -308,8 +309,9 @@ def download_deps(self):
         self.download("datasets", "https://raw.githubusercontent.com/fgnt/mnist/master/t10k-labels-idx1-ubyte.gz", "t10k-labels.idx1-ubyte.gz", unzip=True, skip_data_dir=True)
 
     def extra_cmake_args(self):
+        oneapi = get_oneapi()
         return [
-            f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{self.oneapi.dnn_include()} -I{self.oneapi.mkl_include()} -L{self.oneapi.dnn_lib()} -L{self.oneapi.mkl_lib()}"
+            f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}"
         ]
 
     def name(self):
@@ -337,15 +339,15 @@ def parse_output(self, stdout: str) -> float:
 
 class SVM(VelocityBase):
     def __init__(self, vb: VelocityBench):
-        self.oneapi = get_oneapi()
         super().__init__("svm", "svm_sycl", vb, "s")
 
     def ld_libraries(self):
-        return self.oneapi.ld_libraries()
+        return get_oneapi().ld_libraries()
 
     def extra_cmake_args(self):
+        oneapi = get_oneapi()
         return [
-            f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{self.oneapi.dnn_include()} -I{self.oneapi.mkl_include()} -L{self.oneapi.dnn_lib()} -L{self.oneapi.mkl_lib()}"
+            f"-DCMAKE_CXX_FLAGS=-O3 -fsycl -ffast-math -I{oneapi.dnn_include()} -I{oneapi.mkl_include()} -L{oneapi.dnn_lib()} -L{oneapi.mkl_lib()}"
         ]
 
     def name(self):

scripts/benchmarks/history.py

@@ -7,7 +7,7 @@
 import json
 from pathlib import Path
 from benches.result import Result, BenchmarkRun
-from benches.options import Compare, options
+from options import Compare, options
 from datetime import datetime, timezone
 from utils.utils import run;
 

scripts/benchmarks/main.py

@@ -11,11 +11,12 @@
 from benches.llamacpp import *
 from benches.umf import *
 from benches.test import TestSuite
-from benches.options import Compare, options
+from options import Compare, options
 from output_markdown import generate_markdown
 from output_html import generate_html
 from history import BenchmarkHistory
-from utils.utils import prepare_workdir;
+from utils.utils import prepare_workdir
+from utils.compute_runtime import *
 
 import argparse
 import re
@@ -117,6 +118,13 @@ def process_results(results: dict[str, list[Result]], stddev_threshold_override)
 def main(directory, additional_env_vars, save_name, compare_names, filter):
     prepare_workdir(directory, INTERNAL_WORKDIR_VERSION)
 
+    if options.build_compute_runtime:
+        print(f"Setting up Compute Runtime {options.compute_runtime_tag}")
+        cr = get_compute_runtime()
+        print("Compute Runtime setup complete.")
+        options.extra_ld_libraries.extend(cr.ld_libraries())
+        options.extra_env_vars.update(cr.env_vars())
+
     suites = [
         ComputeBench(directory),
         VelocityBench(directory),
@@ -129,15 +137,15 @@ def main(directory, additional_env_vars, save_name, compare_names, filter):
     benchmarks = []
 
     for s in suites:
-        print(f"Setting up {type(s).__name__}")
-        s.setup()
-        print(f"{type(s).__name__} setup complete.")
-
-    for s in suites:
-        benchmarks += s.benchmarks()
+        suite_benchmarks = s.benchmarks()
+        if filter:
+            suite_benchmarks = [benchmark for benchmark in suite_benchmarks if filter.search(benchmark.name())]
 
-    if filter:
-        benchmarks = [benchmark for benchmark in benchmarks if filter.search(benchmark.name())]
+        if suite_benchmarks:
+            print(f"Setting up {type(s).__name__}")
+            s.setup()
+            print(f"{type(s).__name__} setup complete.")
+            benchmarks += suite_benchmarks
 
     for b in benchmarks:
         print(b.name())
@@ -241,7 +249,7 @@ def validate_and_parse_env_args(env_args):
     parser.add_argument("--save", type=str, help='Save the results for comparison under a specified name.')
     parser.add_argument("--compare", type=str, help='Compare results against previously saved data.', action="append", default=["baseline"])
     parser.add_argument("--iterations", type=int, help='Number of times to run each benchmark to select a median value.', default=options.iterations)
-    parser.add_argument("--stddev-threshold", type=float, help='If stddev % is above this threshold, rerun all iterations', default=options.stddev_threshold)
+    parser.add_argument("--stddev-threshold", type=float, help='If stddev pct is above this threshold, rerun all iterations', default=options.stddev_threshold)
     parser.add_argument("--timeout", type=int, help='Timeout for individual benchmarks in seconds.', default=options.timeout)
     parser.add_argument("--filter", type=str, help='Regex pattern to filter benchmarks by name.', default=None)
     parser.add_argument("--epsilon", type=float, help='Threshold to consider change of performance significant', default=options.epsilon)
@@ -252,12 +260,8 @@ def validate_and_parse_env_args(env_args):
     parser.add_argument("--output-html", help='Create HTML output', action="store_true", default=False)
     parser.add_argument("--output-markdown", help='Create Markdown output', action="store_true", default=True)
     parser.add_argument("--dry-run", help='Do not run any actual benchmarks', action="store_true", default=False)
-    parser.add_argument(
-    "--iterations-stddev",
-    type=int,
-    help="Max number of iterations of the loop calculating stddev after completed benchmark runs",
-    default=options.iterations_stddev,
-    )
+    parser.add_argument("--compute-runtime", nargs='?', const=options.compute_runtime_tag, help="Fetch and build compute runtime")
+    parser.add_argument("--iterations-stddev", type=int, help="Max number of iterations of the loop calculating stddev after completed benchmark runs", default=options.iterations_stddev)
 
     args = parser.parse_args()
     additional_env_vars = validate_and_parse_env_args(args.env)
@@ -279,6 +283,9 @@ def validate_and_parse_env_args(env_args):
     options.dry_run = args.dry_run
     options.umf = args.umf
     options.iterations_stddev = args.iterations_stddev
+    if args.compute_runtime is not None:
+        options.build_compute_runtime = True
+        options.compute_runtime_tag = args.compute_runtime
 
     benchmark_filter = re.compile(args.filter) if args.filter else None
 

scripts/benchmarks/benches/options.py renamed to scripts/benchmarks/options.py

@@ -1,4 +1,4 @@
-from dataclasses import dataclass
+from dataclasses import dataclass, field
 from enum import Enum
 
 class Compare(Enum):
@@ -27,6 +27,10 @@ class Options:
     stddev_threshold: float = 0.02
     epsilon: float = 0.02
     iterations_stddev: int = 5
+    build_compute_runtime: bool = False
+    extra_ld_libraries: list[str] = field(default_factory=list)
+    extra_env_vars: dict = field(default_factory=dict)
+    compute_runtime_tag: str = 'c1ed0334d65f6ce86d7273fe4137d1d4a5b5fa7c'
 
 options = Options()
 

scripts/benchmarks/output_markdown.py

@@ -5,7 +5,7 @@
 
 import collections, re
 from benches.result import Result
-from benches.options import options
+from options import options
 import math
 
 class OutputLine: