pfm parsing for perf stat codes per instruction

.gitignore

@@ -5,4 +5,6 @@ bin
 __pycache__
 dist
 phlop.egg-info/
-scope_timer.txt
+
+*scope_timer.txt
+tpp

mkn.yaml

@@ -7,6 +7,28 @@ profile:
 - name: base
   inc: inc
 
+- name: pfm
+  inc: tpp/pfm/include
+  sub: pfm&tpp/pfm(git://perfmon2.git.sourceforge.net/gitroot/perfmon2/libpfm4)
+
+- name: pfm_lib
+  parent: pfm
+  mode: static
+  inc: |
+    tpp/pfm/include
+    tpp/pfm/lib, 0
+  src: tpp/pfm/lib
+
+- name: pfm_events
+  self: pfm_lib
+  main: tpp/pfm/examples/check_events.c
+  out: check_events
+
+- name: pfm_info
+  self: pfm_lib
+  main: tpp/pfm/examples/showevtinfo.c
+  out: showevtinfo
+
 - name: scope_timer
   parent: base
   src: src/phlop/timing/scope_timer.cpp
@@ -18,3 +40,5 @@ profile:
   src: src/phlop/timing/threaded_scope_timer.cpp
   mode: shared
   test: tests/timing/test_threaded_scope_timer.cpp
+
+

phlop/app/perf.py

@@ -117,6 +117,7 @@ def cli_args_parser(description="Perf tool"):
         logging="0=off, 1=on non zero exit code, 2=always",
         outfile="path for saved file if active",
         tool="stat/record/etc",
+        extra="forward string to perf command",
     )
 
     parser = argparse.ArgumentParser(
@@ -133,6 +134,8 @@ def cli_args_parser(description="Perf tool"):
     parser.add_argument("-o", "--outfile", default=None, help=_help.outfile)
     parser.add_argument("-t", "--tool", default="stat", help=_help.tool)
     parser.add_argument("--logging", type=int, default=1, help=_help.logging)
+    parser.add_argument("-e", "--extra", type=str, default="", help=_help.extra)
+
     return parser
 
 

phlop/app/pfm/__init__.py

@@ -0,0 +1,5 @@
+#
+#
+#
+#
+#

phlop/app/pfm/check_events.py

@@ -0,0 +1,48 @@
+#
+#
+#
+#
+#
+
+
+import logging
+from pathlib import Path
+
+from phlop.os import pushd
+from phlop.proc import run
+from phlop.string import decode_bytes
+
+FILE_DIR = Path(__file__).resolve().parent
+
+logger = logging.getLogger(__name__)
+check_events_start = "Total events:"
+
+
+def parse_check_events_output(lines):
+    return lines[-1].split(":")[1].strip().replace("0x", "r")
+
+
+def run_check_events(code):
+    with pushd(FILE_DIR.parent.parent.parent):
+        return decode_bytes(
+            run(f"./tpp/pfm/examples/check_events {code}").stdout
+        ).splitlines()
+
+
+def get_evt_perf_code(code):
+    return parse_check_events_output(run_check_events(code))
+
+
+if __name__ == "__main__":
+    from phlop.app.pfm.showevtinfo import get_evt_info
+
+    key, code = "[MULT_FLOPS]", ""
+    for info in get_evt_info():
+        if key in info.umask:
+            code = f"{info.name}:{info.umask[key].code}"
+            break
+
+    assert code != ""
+
+    # print("get_evt_perf_code", get_evt_perf_code(code))
+    print(run(f"perf stat -e {get_evt_perf_code(code)} sleep 5"))

phlop/app/pfm/showevtinfo.py

@@ -0,0 +1,81 @@
+#
+#
+#
+#
+#
+
+
+import logging
+from dataclasses import dataclass, field
+from pathlib import Path
+
+from phlop.os import pushd
+from phlop.proc import run
+from phlop.string import decode_bytes
+
+FILE_DIR = Path(__file__).resolve().parent
+
+logger = logging.getLogger(__name__)
+EVTINFO_delimiter = "#-----------------------------"
+
+
+@dataclass
+class EVTUMask:
+    id: str
+    desc: str
+    code: str
+
+
+@dataclass
+class EVTInfo:
+    idx: str
+    pmu: str
+    name: str
+    umask: dict = field(default_factory=lambda: {})
+    ect: dict = field(default_factory=lambda: {})
+
+
+def _parse_evtinfo(bits_list):
+    assert len(bits_list) >= 7
+    info = EVTInfo(
+        idx=bits_list[0][1].strip(),
+        pmu=bits_list[1][1].strip(),
+        name=bits_list[2][1].strip(),
+    )
+    for bits in bits_list[7:]:
+        if bits[0].strip().startswith("Umask"):
+            info.umask[bits[3].strip()] = EVTUMask(
+                id=bits[3].strip(), desc=bits[5].strip(), code=bits[1].strip()
+            )
+    return info
+
+
+def parse_evtinfo_output(lines):
+    start_idx = 0
+    for line in lines:
+        start_idx += 1
+        if line.strip() == EVTINFO_delimiter:
+            break
+
+    bits_list, results = [], []
+    for line in lines[start_idx:]:
+        if line == EVTINFO_delimiter:
+            results.append(_parse_evtinfo(bits_list))
+            bits_list = []
+            continue
+        bits_list.append(line.strip().split(":"))
+
+    return results
+
+
+def run_evtinfo():
+    with pushd(FILE_DIR.parent.parent.parent):
+        return decode_bytes(run("./tpp/pfm/examples/showevtinfo").stdout).splitlines()
+
+
+def get_evt_info():
+    return parse_evtinfo_output(run_evtinfo())
+
+
+if __name__ == "__main__":
+    print(get_evt_info())

phlop/app/stats_man.py

@@ -20,7 +20,6 @@
 from phlop.dict import ValDict
 from phlop.proc import run_raw
 
-logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger(__name__)
 
 _default_interval = 2

phlop/run/mpirun_perf.py

@@ -7,7 +7,6 @@
 
 from phlop.app import perf as p
 
-logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger(__name__)
 
 MPI_RANK = os.environ.get("OMPI_COMM_WORLD_RANK")

phlop/run/mpirun_stats_man.py

@@ -7,7 +7,6 @@
 
 from phlop.app import stats_man as sman
 
-logging.basicConfig(level=logging.DEBUG)
 logger = logging.getLogger(__name__)
 
 MPI_RANK = os.environ.get("OMPI_COMM_WORLD_RANK")

phlop/run/perf.py

@@ -19,11 +19,11 @@
 """
 
 
-def perf_stat_cmd(cli_args, path, line):
+def perf_stat_cmd(cli_args, path, line, options):
     file = Path(line.split(" ")[-1]).stem
     outpath = logpath / path.stem
     outpath.mkdir(parents=True, exist_ok=True)
-    return p.stat_cmd(line, p.stat_events, outpath / f"{file}.json")
+    return p.stat_cmd(line, p.stat_events, outpath / f"{file}.json", options)
 
 
 def get_from_files(cli_args):
@@ -50,7 +50,11 @@ def get_remaining(cli_args):
     test_batches = {}
     path = Path(cli_args.remaining[-1]).parent
     test_case = tc.determine_cores_for_test_case(
-        tc.TestCase(cmd=perf_stat_cmd(cli_args, path, " ".join(cli_args.remaining)))
+        tc.TestCase(
+            cmd=perf_stat_cmd(
+                cli_args, path, " ".join(cli_args.remaining), cli_args.extra
+            )
+        )
     )
     test_batches[test_case.cores] = [test_case]
     return [tc.TestBatch(v, k) for k, v in test_batches.items()]

phlop/testing/parallel_processor.py

@@ -16,8 +16,7 @@
 logger = getLogger(__name__)
 
 
-class TestCaseFailure(Exception):
-    ...
+class TestCaseFailure(Exception): ...
 
 
 class LoggingMode(Enum):

sh/setup_pfm.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+
+CWD="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" && cd "$CWD"/..
+
+set -ex
+
+[ ! -d "tpp/pfm" ] && (
+    git clone git://perfmon2.git.sourceforge.net/gitroot/perfmon2/libpfm4 tpp/pfm
+    cd tpp/pfm
+    make
+)

sh/test.sh

@@ -1,7 +1,6 @@
 #!/usr/bin/env bash
 
-CWD="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
-cd "$CWD"/..
+CWD="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )" && cd "$CWD"/..
 
 set -ex
 
@@ -19,4 +18,7 @@ py -O tests/timing/test_scope_timer.py test_scope_timer -f scope_timer.txt
 py -Om phlop.run.valgrind echo yes
 py -Om phlop.run.valgrind --tool=massif echo yes
 
-py -Om phlop.run.perf echo yes || echo "perf failed, assumed CI"
+py -Om phlop.run.perf -e="--all-user" echo yes || echo "perf failed, assumed CI"
+
+# install via ./sh/setup_pfm.sh
+[ -d "tpp/pfm" ] && py -O tests/_phlop/app/pfm/test_pfm.py || echo "pfm missing, skipped"

tests/_phlop/app/pfm/test_pfm.py

@@ -0,0 +1,27 @@
+#
+#
+#
+
+
+from phlop.proc import run
+from phlop.string import decode_bytes
+
+if __name__ == "__main__":
+    from phlop.app.pfm.check_events import get_evt_perf_code
+    from phlop.app.pfm.showevtinfo import get_evt_info
+
+    code = ""
+    key0, key1 = "[MULT_FLOPS]", "[ADD_SUB_FLOPS]"
+    for info in get_evt_info():
+        if key0 in info.umask:
+            for key, umask in info.umask.items():
+                code += f"{info.name}:{umask.code} "
+            break
+        # if key1 in info.umask:
+        #     code += f"{info.name}:{info.umask[key1].code} "
+
+    code = code.strip()
+    assert code != ""
+
+    events = " ".join([f"-e {get_evt_perf_code(ev)}" for ev in code.split(" ")])
+    print(decode_bytes(run(f"perf stat {events} sleep 5").stderr))