bench now compares aot

This commit is contained in:
2026-02-17 13:42:36 -06:00
parent ad26e71ad1
commit 78051e24f3

142
bench.ce
View File

@@ -13,6 +13,25 @@ var _args = args == null ? [] : args
var target_pkg = null // null = current package
var target_bench = null // null = all benchmarks, otherwise specific bench file
var all_pkgs = false
var bench_mode = "bytecode" // "bytecode", "native", or "compare"
// Strip mode flags from args before parsing
function strip_mode_flags() {
var filtered = []
arrfor(_args, function(a) {
if (a == '--native') {
bench_mode = "native"
} else if (a == '--bytecode') {
bench_mode = "bytecode"
} else if (a == '--compare') {
bench_mode = "compare"
} else {
push(filtered, a)
}
})
_args = filtered
}
strip_mode_flags()
// Benchmark configuration
def WARMUP_BATCHES = 3
@@ -394,6 +413,45 @@ function format_ops(ops) {
return `${round(ops / 1000000000 * 100) / 100}G ops/s`
}
// Load a module for benchmarking in the given mode
// Returns the module value, or null on failure
function load_bench_module(f, package_name, mode) {
var mod_path = text(f, 0, -3)
var use_pkg = package_name ? package_name : fd.realpath('.')
var prefix = null
var src_path = null
if (mode == "native") {
prefix = testlib.get_pkg_dir(package_name)
src_path = prefix + '/' + f
return shop.use_native(src_path, use_pkg)
}
return shop.use(mod_path, use_pkg)
}
// Collect benchmark functions from a loaded module
function collect_bench_fns(bench_mod) {
var benches = []
if (is_function(bench_mod)) {
push(benches, {name: 'main', fn: bench_mod})
} else if (is_object(bench_mod)) {
arrfor(array(bench_mod), function(k) {
if (is_function(bench_mod[k]))
push(benches, {name: k, fn: bench_mod[k]})
})
}
return benches
}
// Print results for a single benchmark
function print_bench_result(result, label) {
var prefix = label ? `[${label}] ` : ''
log.console(` ${prefix}${format_ns(result.median_ns)}/op ${format_ops(result.ops_per_sec)}`)
log.console(` ${prefix}min: ${format_ns(result.min_ns)} max: ${format_ns(result.max_ns)} stddev: ${format_ns(result.stddev_ns)}`)
if (result.batch_size > 1) {
log.console(` ${prefix}batch: ${result.batch_size} samples: ${result.samples}`)
}
}
// Run benchmarks for a package
function run_benchmarks(package_name, specific_bench) {
var bench_files = collect_benches(package_name, specific_bench)
@@ -406,15 +464,16 @@ function run_benchmarks(package_name, specific_bench) {
if (length(bench_files) == 0) return pkg_result
if (package_name) log.console(`Running benchmarks for ${package_name}`)
else log.console(`Running benchmarks for local package`)
var mode_label = bench_mode == "compare" ? "bytecode vs native" : bench_mode
if (package_name) log.console(`Running benchmarks for ${package_name} (${mode_label})`)
else log.console(`Running benchmarks for local package (${mode_label})`)
arrfor(bench_files, function(f) {
var mod_path = text(f, 0, -3)
var load_error = false
var bench_mod = null
var use_pkg = null
var benches = []
var native_benches = []
var bench_mod = null
var native_mod = null
var error_result = null
var file_result = {
@@ -423,16 +482,21 @@ function run_benchmarks(package_name, specific_bench) {
}
var _load_file = function() {
use_pkg = package_name ? package_name : fd.realpath('.')
bench_mod = shop.use(mod_path, use_pkg)
if (is_function(bench_mod)) {
push(benches, {name: 'main', fn: bench_mod})
} else if (is_object(bench_mod)) {
arrfor(array(bench_mod), function(k) {
if (is_function(bench_mod[k]))
push(benches, {name: k, fn: bench_mod[k]})
})
var _load_native = null
if (bench_mode == "compare") {
bench_mod = load_bench_module(f, package_name, "bytecode")
benches = collect_bench_fns(bench_mod)
_load_native = function() {
native_mod = load_bench_module(f, package_name, "native")
native_benches = collect_bench_fns(native_mod)
} disruption {
log.console(` ${f}: native compilation failed, comparing skipped`)
native_benches = []
}
_load_native()
} else {
bench_mod = load_bench_module(f, package_name, bench_mode)
benches = collect_bench_fns(bench_mod)
}
if (length(benches) > 0) {
@@ -440,18 +504,50 @@ function run_benchmarks(package_name, specific_bench) {
arrfor(benches, function(b) {
var bench_error = false
var result = null
var nat_b = null
var nat_error = false
var nat_result = null
var _run_bench = function() {
var speedup = 0
var _run_nat = null
result = run_single_bench(b.fn, b.name)
result.package = pkg_result.package
result.mode = bench_mode == "compare" ? "bytecode" : bench_mode
push(file_result.benchmarks, result)
pkg_result.total++
log.console(` ${result.name}`)
log.console(` ${format_ns(result.median_ns)}/op ${format_ops(result.ops_per_sec)}`)
log.console(` min: ${format_ns(result.min_ns)} max: ${format_ns(result.max_ns)} stddev: ${format_ns(result.stddev_ns)}`)
if (result.batch_size > 1) {
log.console(` batch: ${result.batch_size} samples: ${result.samples}`)
if (bench_mode == "compare") {
print_bench_result(result, "bytecode")
// Find matching native bench and run it
nat_b = find(native_benches, function(nb) { return nb.name == b.name })
if (nat_b) {
_run_nat = function() {
nat_result = run_single_bench(nat_b.fn, b.name)
nat_result.package = pkg_result.package
nat_result.mode = "native"
push(file_result.benchmarks, nat_result)
pkg_result.total++
print_bench_result(nat_result, "native ")
if (nat_result.median_ns > 0) {
speedup = result.median_ns / nat_result.median_ns
log.console(` speedup: ${round(speedup * 100) / 100}x`)
}
} disruption {
nat_error = true
}
_run_nat()
if (nat_error) {
log.console(` [native ] ERROR`)
}
} else {
log.console(` [native ] (no matching function)`)
}
} else {
print_bench_result(result, null)
}
} disruption {
bench_error = true
@@ -524,8 +620,10 @@ function generate_reports() {
var report_dir = shop.get_reports_dir() + '/bench_' + timestamp
testlib.ensure_dir(report_dir)
var mode_str = bench_mode == "compare" ? "bytecode vs native" : bench_mode
var txt_report = `BENCHMARK REPORT
Date: ${time.text(time.number())}
Mode: ${mode_str}
Total benchmarks: ${total_benches}
=== SUMMARY ===
@@ -536,10 +634,11 @@ Total benchmarks: ${total_benches}
arrfor(pkg_res.files, function(f) {
txt_report += ` ${f.name}\n`
arrfor(f.benchmarks, function(b) {
var mode_tag = b.mode ? ` [${b.mode}]` : ''
if (b.error) {
txt_report += ` ERROR ${b.name}: ${b.error}\n`
} else {
txt_report += ` ${b.name}: ${format_ns(b.median_ns)}/op (${format_ops(b.ops_per_sec)})\n`
txt_report += ` ${b.name}${mode_tag}: ${format_ns(b.median_ns)}/op (${format_ops(b.ops_per_sec)})\n`
}
})
})
@@ -553,7 +652,8 @@ Total benchmarks: ${total_benches}
arrfor(f.benchmarks, function(b) {
if (b.error) return
txt_report += `\n${pkg_res.package}::${b.name}\n`
var detail_mode = b.mode ? ` [${b.mode}]` : ''
txt_report += `\n${pkg_res.package}::${b.name}${detail_mode}\n`
txt_report += ` batch_size: ${b.batch_size} samples: ${b.samples}\n`
txt_report += ` median: ${format_ns(b.median_ns)}/op\n`
txt_report += ` mean: ${format_ns(b.mean_ns)}/op\n`