mirror of
https://github.com/HeyPuter/gex.git
synced 2026-11-13 20:12:43 -10:00
velzie
f7ed326302
Groundwork for running Debian's own glxgears in the page, against Debian's own Mesa. Three pieces, none of which is the demo yet. An 81 MB rootfs, from 342. `web/build-graphics-rootfs.sh` takes the closure a traced run actually opens -- 204 MB -- and then removes the two libraries that are most of it. Debian ships one megadriver: `swrast_dri.so` is every Gallium driver at once and declares libLLVM-15 in DT_NEEDED because llvmpipe wants it, 106 MB, plus 21 MB of libz3 that LLVM in turn wants. The virgl path touches none of it, but ld.so maps it before a line of driver code runs. So the image carries a stand-in: the 280 symbols the megadriver imports from LLVM, each one a `brk #0`. 127 MB removed for 40 KB added, and if anything ever does call in it dies at the instruction rather than drawing something subtly wrong. The stub has to carry LLVM's symbol version. Unversioned, glibc's ld.so does not warn -- it fails an assertion in dl-lookup.c and kills the process before main. A canvas. vtest has no present, and that is what vtest is: the client reads its pixels back with TRANSFER_GET and compares them. An X server is happy never to see a screen. So every readback of a colour surface is now also blitted to the canvas, and `gex_virgl_gl_state` counts the blits -- because a run that renders perfectly and shows nothing looks exactly like one that does not render, and until now there was no way to tell them apart. And a test that reports over HTTP instead of being dumped. `--virtual-time-budget` advances the clock for every timer, so a five-second sampler burned 1,750 seconds of a half-hour budget in two minutes of real time and the browser was killed with the X server still starting. The page posts its progress and its result; the run takes as long as it takes.
83 lines
3.1 KiB
Bash
Executable File
83 lines
3.1 KiB
Bash
Executable File
#!/bin/sh
|
|
# One nbench kernel, repeated, in the browser runtime. Prints the median.
|
|
#
|
|
# tests/real-programs/kernel.sh <KERNEL> <module.wasm> [runs] [seconds]
|
|
#
|
|
# The full `nbench.sh` table cannot resolve anything small. Two seconds a
|
|
# kernel on a shared machine gives about 30% spread -- QEMU running the same
|
|
# binary three times gave 8.82, 11.19 and 11.46 on ASSIGNMENT -- so a change
|
|
# worth 10% is invisible in it, and a change worth 10% is most changes.
|
|
#
|
|
# This runs one kernel at a time, for longer, several times over, and reports
|
|
# the median with the spread beside it so the reader can see whether the
|
|
# number deserves any confidence. Comparing two modules means running this
|
|
# twice and looking at whether the ranges overlap.
|
|
#
|
|
# KERNEL is nbench's own name with spaces removed: NUMERICSORT, STRINGSORT,
|
|
# BITFIELD, FPEMULATION, FOURIER, ASSIGNMENT, IDEA, HUFFMAN, NEURALNET,
|
|
# LUDECOMPOSITION.
|
|
set -eu
|
|
|
|
KERNEL="${1:?usage: kernel.sh <KERNEL> <module.wasm> [runs] [seconds]}"
|
|
WASM="${2:?usage: kernel.sh <KERNEL> <module.wasm> [runs] [seconds]}"
|
|
RUNS="${3:-5}"
|
|
SECONDS_PER="${4:-4}"
|
|
|
|
HERE="$(cd "$(dirname "$0")" && pwd)"
|
|
ROOT="$(cd "$HERE/../.." && pwd)"
|
|
WORK="${GEX_NBENCH_DIR:-${TMPDIR:-/tmp}/gex-nbench}"
|
|
D="$WORK/nbench-byte-2.2.3"
|
|
|
|
[ -x "$D/nbench" ] || { echo "skip: no nbench at $D (run nbench.sh once)"; exit 0; }
|
|
[ -f "$WASM" ] || { echo "skip: no module at $WASM"; exit 0; }
|
|
command -v node >/dev/null 2>&1 || { echo "skip: no node"; exit 0; }
|
|
|
|
# nbench's config turns kernels off by name; leaving one on is what makes
|
|
# this a measurement of that kernel rather than of the whole suite.
|
|
cfg="$D/ONLY.$$"
|
|
{
|
|
echo "MINSECONDS=$SECONDS_PER"
|
|
echo "ALLSTATS=F"
|
|
for k in NUMSORT STRINGSORT BITFIELD EMF FOUR ASSIGN IDEA HUFF NNET LU; do
|
|
echo "DO$k=F"
|
|
done
|
|
} >"$cfg"
|
|
# The config name nbench wants is the flag, not the kernel: its own switches
|
|
# are abbreviated and do not match the printed titles.
|
|
case "$KERNEL" in
|
|
NUMERICSORT) flag=NUMSORT ;;
|
|
STRINGSORT) flag=STRINGSORT ;;
|
|
BITFIELD) flag=BITFIELD ;;
|
|
FPEMULATION) flag=EMF ;;
|
|
FOURIER) flag=FOUR ;;
|
|
ASSIGNMENT) flag=ASSIGN ;;
|
|
IDEA) flag=IDEA ;;
|
|
HUFFMAN) flag=HUFF ;;
|
|
NEURALNET) flag=NNET ;;
|
|
LUDECOMPOSITION) flag=LU ;;
|
|
*) echo "unknown kernel $KERNEL"; exit 1 ;;
|
|
esac
|
|
sed -i "s/^DO$flag=F$/DO$flag=T/" "$cfg"
|
|
|
|
rates=""
|
|
i=0
|
|
while [ "$i" -lt "$RUNS" ]; do
|
|
i=$((i + 1))
|
|
r=$(node "$ROOT/shims/browser/js/run-node.mjs" \
|
|
--dir "$D=$D" --cwd "$D" --argv0 "$D/nbench" \
|
|
"$WASM" "$D/nbench" "-c$(basename "$cfg")" 2>/dev/null |
|
|
awk -F: '/^ +: *[0-9]/ { gsub(/ /, "", $2); if ($2 + 0 > 0) { print $2; exit } }')
|
|
[ -n "$r" ] || { echo "run $i produced nothing"; continue; }
|
|
rates="$rates $r"
|
|
done
|
|
rm -f "$cfg"
|
|
|
|
echo "$rates" | tr ' ' '\n' | grep -v '^$' | sort -g | awk -v k="$KERNEL" '
|
|
{ v[NR] = $1 }
|
|
END {
|
|
if (NR == 0) { print "no results"; exit 1 }
|
|
med = (NR % 2) ? v[(NR + 1) / 2] : (v[NR / 2] + v[NR / 2 + 1]) / 2
|
|
printf "%-16s median %-12g min %-12g max %-12g spread %.1f%% (n=%d)\n",
|
|
k, med, v[1], v[NR], 100 * (v[NR] - v[1]) / med, NR
|
|
}'
|