[Tools] (P14): the sampler and profiler find spawn and tcp server processes by their install path, plus a per-component profile split.

tools/device_bench/disagg/components.py

@@ -0,0 +1,66 @@
+#!/usr/bin/env python3
+"""Inclusive share of fixed components on one thread of a simpleperf profile, converted to
+milliseconds per frame with the thread's measured CPU ms/frame.
+
+    python components.py <perf.data> <binary_cache> <thread comm> [cpu_ms_per_frame]
+
+A sample counts toward a component when any frame of its call chain matches the component's
+pattern (first match in COMPONENTS order wins for the 'exclusive' column, so the rows of that
+column add up to 100%).
+"""
+import os
+import sys
+from collections import Counter
+
+sys.path.insert(0, os.environ.get("SIMPLEPERF_DIR", ""))
+from simpleperf_report_lib import ReportLib  # noqa: E402
+
+COMPONENTS = [
+    ("trace parse (harness)", ("trace::Parser::", "BrotliFile", "trace::Call::~Call", "trace::File::")),
+    ("validate point", ("MGPipeValidateForVerb",)),
+    ("backend draw setup", ("SetupWireDraw", "VulkanRenderer::SetupDraw", "DirectGLES::PrepareForDraw")),
+    ("backend draw other", ("DirectVulkan::DrawElements", "DirectGLES::DrawElements", "VulkanRenderer::DrawElements")),
+    ("verb port / record sink", ("EmitDrawRecord", "VerbChannel::Port", "RecordVerbSink::")),
+    ("wire encode (client)", ("ClientSession::Emit", "PipeWireEncoder::")),
+    ("apply decode", ("PipeWireDecoder::", "PipeApplier::ApplyOne")),
+    ("apply loop idle/poll", ("ServerLoop::ApplyThreadMain", "ServerLoop::DrainRing")),
+    ("other GL entry points", ("retrace::retraceCall",)),
+]
+
+
+def main():
+    data, symfs, comm = sys.argv[1], sys.argv[2], sys.argv[3]
+    cpu = float(sys.argv[4]) if len(sys.argv) > 4 else None
+    lib = ReportLib()
+    lib.SetRecordFile(data)
+    lib.SetSymfs(symfs)
+    total = 0
+    inclusive = Counter()
+    exclusive = Counter()
+    while True:
+        sample = lib.GetNextSample()
+        if sample is None:
+            break
+        if comm not in sample.thread_comm:
+            continue
+        total += 1
+        frames = [lib.GetSymbolOfCurrentSample().symbol_name]
+        chain = lib.GetCallChainOfCurrentSample()
+        frames += [chain.entries[i].symbol.symbol_name for i in range(chain.nr)]
+        first = None
+        for name, patterns in COMPONENTS:
+            if any(p in f for f in frames for p in patterns):
+                inclusive[name] += 1
+                if first is None:
+                    first = name
+        exclusive[first or "rest"] += 1
+    print(f"{comm}: {total} samples" + (f", {cpu:.2f} ms/frame" if cpu else ""))
+    for name, _ in COMPONENTS + [("rest", ())]:
+        inc = 100.0 * inclusive[name] / max(1, total)
+        exc = 100.0 * exclusive[name] / max(1, total)
+        ms = f"  {cpu * exc / 100.0:6.2f} ms" if cpu else ""
+        print(f"  {name:28} incl {inc:5.1f}%  excl {exc:5.1f}%{ms}")
+
+
+if __name__ == "__main__":
+    main()

tools/device_bench/disagg/cpusampler.c

@@ -95,9 +95,9 @@ static int ScanPids(const char* pkg, int* pids) {
         snprintf(path, sizeof path, "/proc/%s/cmdline", e->d_name);
         if (ReadSmall(path, buf, sizeof buf) <= 0) continue;
         // The replay process (and its :services), plus the render-server processes the spawn and
-        // tcp arms exec (their cmdline is the server library's name, not the package's).
+        // tcp arms exec (their cmdline is the server library's path in the package's install dir).
         if ((strncmp(buf, pkg, len) == 0 && (buf[len] == 0 || buf[len] == ':')) ||
-            strncmp(buf, "libMobileGLServer.so", 20) == 0)
+            (strstr(buf, "libMobileGLServer.so") != NULL && strstr(buf, pkg) != NULL))
             pids[n++] = atoi(e->d_name);
     }
     closedir(d);

tools/device_bench/disagg/profile_p14.sh

@@ -33,7 +33,7 @@ esac
 adb -s "$SER" shell "am force-stop $PKG_FEAT; am force-stop $PKG_DEV; run-as $pkg sh -c 'rm -f $app/output/*.json'" >/dev/null 2>&1
 adb -s "$SER" shell "am start -a top.mobilegl.plugin.TRACE_REPLAY -n $pkg/top.mobilegl.plugin.trace.TraceReplayActivity --es trace_path $app/input-$WL/trace.trace --es golden_path $app/input-$WL/golden.png --es output_dir $app/output --es backend $BACKEND $args --ez use_pbuffer true --ez benchmark true --es benchmark_result_path $app/output/benchmark.json --es mobilegl_env '$env'" >/dev/null
 sleep "$DELAY"
-pids=$(adb -s "$SER" shell "su -c 'echo \$(pidof $pkg) \$(pgrep -f ^libMobileGLServer.so)'" | tr -d '\r' | tr -s ' ' ',' | sed 's/^,//; s/,$//')
+pids=$(adb -s "$SER" shell "su -c 'echo \$(pidof $pkg) \$(pgrep -f $pkg.*libMobileGLServer.so)'" | tr -d '\r' | tr -s ' ' ',' | sed 's/^,//; s/,$//')
 echo "pids=$pids" | tee "$OUT/pids.txt"
 adb -s "$SER" shell "su -c 'simpleperf record -e cpu-clock -p $pids --call-graph fp -f 4000 --duration $DUR -o /data/local/tmp/p14perf.data'" >"$OUT/record.log" 2>&1
 adb -s "$SER" exec-out "run-as $pkg cat $app/output/benchmark.json" >"$OUT/benchmark.json" 2>/dev/null