instruction mix utility and cleanup (#56)

* remove tsc rounding for more accuracy

* instruction mixes

* revert version

* check for perf

events/clx_skx.txt

@@ -125,7 +125,6 @@ cpu-cycles,
 ref-cycles,
 instructions;
 
-cpu/event=0xb0,umask=0x10,period=100003,name='OFFCORE_REQUESTS.L3_MISS_DEMAND_DATA_RD'/,
 cpu/event=0xd1,umask=0x40,period=100007,name='MEM_LOAD_RETIRED.FB_HIT'/,
 cpu/event=0xa3,umask=0x10,cmask=0x16,period=2000003,name='CYCLE_ACTIVITY.CYCLES_MEM_ANY'/,
 cpu/event=0xa3,umask=0x08,cmask=0x08,period=2000003,name='CYCLE_ACTIVITY.CYCLES_L1D_MISS'/,
@@ -134,7 +133,6 @@ ref-cycles,
 instructions;
 
 cpu/event=0x79,umask=0x30,period=2000003,name='IDQ.MS_UOPS'/,
-cpu/event=0x60,umask=0x10,period=2000003,name='OFFCORE_REQUESTS_OUTSTANDING.L3_MISS_DEMAND_DATA_RD'/,
 cpu/event=0x83,umask=0x04,period=200003,name='ICACHE_64B.IFTAG_STALL'/,
 cpu/event=0x08,umask=0x20,cmask=0x01,period=2000003,name='DTLB_LOAD_MISSES.STLB_HIT:c1'/,
 cpu-cycles,

events/icx.txt

@@ -18,8 +18,6 @@ instructions;
 
 cpu/event=0xf1,umask=0x1f,period=100003,name='L2_LINES_IN.ALL'/,
 cpu/event=0xd1,umask=0x10,period=100021,name='MEM_LOAD_RETIRED.L2_MISS'/,
-cpu/event=0x60,umask=0x10,period=2000003,name='OFFCORE_REQUESTS_OUTSTANDING.L3_MISS_DEMAND_DATA_RD'/,
-cpu/event=0xb0,umask=0x10,period=100003,name='OFFCORE_REQUESTS.L3_MISS_DEMAND_DATA_RD'/,
 cpu-cycles,
 ref-cycles,
 instructions;

events/metric_icx.json

@@ -86,11 +86,6 @@
         "expression": "[L2_RQSTS.CODE_RD_MISS] / [instructions]",
         "expression-txn": "[L2_RQSTS.CODE_RD_MISS] / [TXN]"
     },
-    {
-        "name": "metric_Average LLC data read miss latency (in clks)",
-        "expression": "[OFFCORE_REQUESTS_OUTSTANDING.L3_MISS_DEMAND_DATA_RD] / [OFFCORE_REQUESTS.L3_MISS_DEMAND_DATA_RD]",
-        "origin": "perfspect"
-    },
     {
         "name": "metric_UPI Data transmit BW (MB/sec) (only data)",
         "expression": "([UNC_UPI_TxL_FLITS.ALL_DATA] * (64 / 9.0) / 1000000) / 1"

events/metric_skx_clx.json

@@ -107,11 +107,6 @@
         "expression-txn": "[OCR.ALL_READS.L3_MISS.REMOTE_HIT_FORWARD] / [TXN]",
         "origin": "perfspect"
     },
-    {
-        "name": "metric_Average LLC data read miss latency (in clks)",
-        "expression": "[OFFCORE_REQUESTS_OUTSTANDING.L3_MISS_DEMAND_DATA_RD] / [OFFCORE_REQUESTS.L3_MISS_DEMAND_DATA_RD]",
-        "origin": "perfspect"
-    },
     {
         "name": "metric_Average LLC data read miss latency (in ns)",
         "expression": "(1000000000 * [UNC_CHA_TOR_OCCUPANCY.IA_MISS.0x40433] / [UNC_CHA_TOR_INSERTS.IA_MISS.0x40433]) / ( [UNC_CHA_CLOCKTICKS] / ([CHAS_PER_SOCKET] * [SOCKET_COUNT]) )",

instruction-mix.py

@@ -0,0 +1,99 @@
+#!/usr/bin/env python3
+
+###########################################################################################################
+# Copyright (C) 2021-2023 Intel Corporation
+# SPDX-License-Identifier: BSD-3-Clause
+###########################################################################################################
+
+import logging
+import os
+import platform
+import shlex
+import subprocess
+import sys
+
+from iced_x86 import Decoder, Formatter, FormatterSyntax  # type: ignore
+from src.common import configure_logging, crash
+from src.perf_helpers import get_perf_list
+
+formatter = Formatter(FormatterSyntax.NASM)
+
+
+def get_insn(insn):
+    global formatter
+    insn = bytes.fromhex(insn.replace(" ", ""))
+    insnraw = formatter.format(Decoder(64, insn).decode()).split(" ", 1)
+    insn = insnraw[0]
+    reg = []
+    if len(insnraw) > 1:
+        if "xmm" in insnraw[1]:
+            reg.append("AVX128")
+        if "ymm" in insnraw[1]:
+            reg.append("AVX256")
+        if "zmm" in insnraw[1]:
+            reg.append("AVX512")
+        if len(reg) > 0:
+            insn += ";" + ",".join(reg)
+    return insn
+
+
+if __name__ == "__main__":
+    configure_logging(".")
+
+    if len(sys.argv) != 2 or not sys.argv[1].isdigit():
+        print('usage: "sudo ./instruction-mix 3  # run for 3 seconds"')
+        sys.exit()
+    if os.geteuid() != 0:
+        crash("Must run as root, please re-run")
+    if platform.system() != "Linux":
+        crash("PerfSpect currently supports Linux only")
+    get_perf_list()
+
+    logging.info("collecting...")
+
+    subprocess.run(shlex.split("perf record -a -F 99 sleep " + sys.argv[1]))
+    rawdata = (
+        subprocess.Popen(
+            shlex.split("perf script -F comm,pid,insn"),
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+        )
+        .communicate()[0]
+        .decode()
+        .split("\n")[:-1]
+    )
+
+    logging.info("postprocessing...")
+
+    processmap = {}
+
+    for row in rawdata:
+        sides = row.split("insn:")
+        if len(sides) > 1:
+            insn = get_insn(sides[1])
+            id = sides[0].split()[0] + ";" + insn
+            if id not in processmap:
+                processmap[id] = 0
+            processmap[id] += 1
+
+    # generate freqs
+    col = ""
+    for p in processmap:
+        col += p + " " + str(processmap[p]) + "\n"
+
+    with open("instruction-mix.svg", "w") as f:
+        subprocess.run(
+            shlex.split(
+                os.path.join(
+                    getattr(
+                        sys, "_MEIPASS", os.path.dirname(os.path.abspath(__file__))
+                    ),
+                    "flamegraph.pl",
+                )
+            ),
+            input=col.encode(),
+            stdout=f,
+        )
+
+    os.chmod("instruction-mix.svg", 0o666)
+    logging.info("generated instruction-mix.svg")

perf-collect.py

@@ -432,10 +432,11 @@ def validate_file(fname):
     elif args.app:
         cmd += f" {args.app}"
 
-    perfargs = shlex.split(cmd)
-    validate_perfargs(perfargs)
     if args.verbose:
         logging.info(cmd)
+    perfargs = shlex.split(cmd)
+    validate_perfargs(perfargs)
+
     psi = []
     logging.info("Collection started!")
     start = time.time()

requirements.txt

@@ -1,5 +1,6 @@
 black
 flake8
+iced-x86
 pytype
 simpleeval
 pandas