feat(perf): show GPU render and compute loadHEADmaster

xe exposes no busy percentage, but its DRM fdinfo carries per-engine
cycle counters. vram.sh becomes gpu.sh: the same scan also sums
drm-cycles-rcs/ccs per client and prints them with the GPU's cycle
counter, and Perf.js turns two scans into a percentage per engine.

The GPU card's headline is now the busier engine, with render (games)
and compute (local LLMs) as bars, then VRAM, temperature and fan.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

1 files changed, 65 insertions, 0 deletions

diff --git a/desktop/modules/perf/gpu.sh b/desktop/modules/perf/gpu.sh
new file mode 100755
index 0000000..31f8a6e
--- /dev/null
+++ b/desktop/modules/perf/gpu.sh
@@ -0,0 +1,65 @@
+#!/bin/sh
+#
+# Copyright (C) 2026 Danilo M. <danix@danix.xyz>
+#
+# This program is free software; you can redistribute it and/or modify
+# it under the terms of the GNU General Public License version 2 as
+# published by the Free Software Foundation.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+# GNU General Public License for more details.
+
+# GPU from the fdinfo scan, every xe fd counted once per drm-client-id
+# because a client holds several fds. Prints one line, "vram rcs_busy
+# rcs_clock ccs_busy ccs_clock": VRAM in bytes (drm-resident-vram0), then
+# per engine the busy cycles summed over clients and the GPU's cycle
+# counter, which Perf.js turns into a percentage between two scans. rcs is
+# render (games), ccs compute (local LLMs). Prints nothing when no xe
+# client is found. With a file argument it reads that instead of scanning,
+# for test_gpu.sh.
+#
+# Summed here rather than in JS: a process holding thousands of DRM fds
+# makes the scan a few hundred KB, which is a lot of string churn to parse
+# in QML every 2 s. It was moved here while chasing a GC underflow that
+# turned out to be the sound OSD's createImageData (AGENTS.md), not this.
+#
+# Fds of a repeated client id with different values: the first one awk's
+# for-in reaches counts, where Perf.js took the first in scan order. They
+# share one client, so the values agree in practice.
+
+scan() {
+    if [ -n "$1" ]; then cat "$1"
+    else grep -sH -e '^drm-driver' -e '^drm-client-id' -e '^drm-resident-vram0' \
+        -e '^drm-cycles-[rc]cs' -e '^drm-total-cycles-[rc]cs' \
+        /proc/[0-9]*/fdinfo/* 2>/dev/null
+    fi
+}
+
+scan "$1" | awk -F: '
+    NF >= 3 && $2 ~ /^drm-/ {
+        v = $3
+        for (i = 4; i <= NF; i++) v = v ":" $i
+        sub(/^[ \t]+/, "", v)
+        d[$1, $2] = v
+        fd[$1]
+    }
+    END {
+        for (p in fd) {
+            id = d[p, "drm-client-id"]
+            if (d[p, "drm-driver"] != "xe" || id == "" || (id in seen)) continue
+            if (split(d[p, "drm-resident-vram0"], n, " ") < 1 || n[1] !~ /^[0-9]+$/) continue
+            for (i = 1; i <= 2; i++) {
+                e = i == 1 ? "rcs" : "ccs"
+                busy[e] += d[p, "drm-cycles-" e]
+                c = d[p, "drm-total-cycles-" e] + 0
+                if (c > clock[e]) clock[e] = c
+            }
+            seen[id]
+            u = n[2] == "KiB" ? 1024 : n[2] == "MiB" ? 1048576 : n[2] == "GiB" ? 1073741824 : 1
+            total += n[1] * u
+            any = 1
+        }
+        if (any) printf "%.0f %.0f %.0f %.0f %.0f\n", total, busy["rcs"], clock["rcs"], busy["ccs"], clock["ccs"]
+    }'