pcbjam/tests/apps/standalone/raytrace-threads/raytrace_threads_test.cpp
2026-06-19 09:22:56 +02:00

505 lines
19 KiB
C++

/**
* raytrace_threads_test.cpp
*
* Minimal, fast-building reproduction of the KiCad WASM raytracer threading
* deadlock (kicad/3d-viewer/3d_rendering/raytracing/render_3d_raytrace_base.cpp,
* the post-process passes), PLUS candidate fixes — all in one binary, switchable
* by URL query param so we build ONCE and test every variant.
*
* ------------------------------------------------------------------------------
* THE PATTERN WE'RE REPRODUCING (raytracer post-process, desktop path):
*
* for (i in 0..N) { std::thread t(worker); t.detach(); } // spawn N workers
* while (threadsFinished < N) // ...then WAIT
* std::this_thread::sleep_for(10ms); // by busy-looping
*
* WHY IT DEADLOCKS IN THE BROWSER:
* - std::thread == a Web Worker. Creating one needs a message processed by the
* browser MAIN-THREAD event loop.
* - KiCad already pre-spawned hardware_concurrency() thread-pool workers at
* startup, so the pre-warmed worker slots
* (-sPTHREAD_POOL_SIZE='navigator.hardwareConcurrency') are already full.
* The raytracer's new threads must therefore be spawned ON DEMAND, which
* again needs the event loop.
* - The busy-wait `sleep_for` NEVER returns to the event loop, so those new
* workers are never created -> threadsFinished never advances -> frozen tab.
*
* THE FIX: don't busy-wait — YIELD. emscripten_sleep() (Asyncify) hands control
* back to the event loop each tick, so the pending worker spawns complete and
* the workers actually run on real cores.
*
* TODO(asyncify-nesting) — IMPORTANT CAVEAT this harness does NOT capture: the
* emscripten_sleep variants (B1/B2/m4) pass here but ABORT the real KiCad 3D viewer
* with `Aborted(invalid state: 1)`. The viewer renders inside the wx modal/event-pump,
* which is already mid-Asyncify-unwind, and emscripten_sleep can't nest on that context.
* This program runs from a clean OnInit, so it never nests. The variant actually ported
* to KiCad is B3 (m=5): a persistent pool + sleep_for busy-wait — NO emscripten_sleep, so
* no nesting. That multi-core port is currently PARKED (git -C kicad stash) pending
* research into whether a nestable yield (fibers / emscripten_fiber_swap / JSPI) works.
*
* To reproduce KiCad faithfully we first pre-spawn `park` "parked" workers that
* block on a condition variable (zero CPU) to occupy the pool slots — exactly
* what KiCad's singleton thread pool does. Set ?park=0 to show that WITHOUT that
* pre-existing pool, variant A does NOT deadlock (hwc threads fit in hwc slots).
*
* ------------------------------------------------------------------------------
* URL params (all optional):
* ?m=0 variant A : detached threads + std::this_thread::sleep_for (EXPECT DEADLOCK)
* ?m=1 variant B1 : detached threads + emscripten_sleep yield (FIX, fresh threads)
* ?m=2 variant B2 : persistent pre-warmed pool + emscripten_sleep (FIX, reused threads)
* ?m=3 variant C : serial on the calling thread (current shipped fallback)
* ?park=K parked workers pre-spawned to exhaust the pool (default = hardware_concurrency)
* ?work=K worker threads the pass uses (default = hardware_concurrency)
* ?blocks=K number of work blocks (default 64)
* ?iters=K compute iterations per block (default 4000000; tune so serial ~2-3s)
* ?passes=K how many times to run the pass (default 1; use >1 to show B2 reuse)
*
* Console contract (the Playwright spec asserts on these):
* [RTPOOL] START m=.. park=.. work=.. blocks=.. iters=.. passes=.. hwc=..
* [RTPOOL] PASS done pass=.. workersRan=.. passMs=..
* [RTPOOL] DEADLOCK pass=.. ... (variant A)
* [RTPOOL] SUCCESS mode=.. workersRan=.. totalMs=..
* [RTPOOL] FAIL deadlocked mode=.. totalMs=.. (variant A)
*/
#include "wx/wx.h"
#include <algorithm>
#include <atomic>
#include <chrono>
#include <climits>
#include <cmath>
#include <condition_variable>
#include <cstdarg>
#include <cstdio>
#include <mutex>
#include <thread>
#include <vector>
#ifdef __EMSCRIPTEN__
#include <emscripten/emscripten.h>
#endif
using clock_t_ = std::chrono::steady_clock;
// ----------------------------------------------------------------------------
// logging: print to stdout AND the browser console (Playwright captures console)
// ----------------------------------------------------------------------------
static void rtlog( const char* fmt, ... )
{
char buf[512];
va_list ap;
va_start( ap, fmt );
vsnprintf( buf, sizeof( buf ), fmt, ap );
va_end( ap );
#ifdef __EMSCRIPTEN__
// NOTE: only console.log here. Emscripten already routes printf/stdout to
// console.log, so doing both would duplicate every line.
EM_ASM( { console.log( UTF8ToString( $0 ) ); }, buf );
#else
printf( "%s\n", buf );
#endif
}
// ----------------------------------------------------------------------------
// read the URL query params in one shot into out[6] = {m,park,work,blocks,iters,passes}.
// INT_MIN means "param absent -> use the C++ default". We write through HEAP32 (an
// exported runtime method) instead of UTF8ToString/setValue, which are NOT guaranteed
// to be present in the EM_ASM scope under this build's EXPORTED_RUNTIME_METHODS.
// ----------------------------------------------------------------------------
static void readUrlParams( int* out /* [6] */ )
{
for( int i = 0; i < 6; ++i )
out[i] = INT_MIN;
#ifdef __EMSCRIPTEN__
EM_ASM(
{
// serve-handler defaults cleanUrls:true, which 301-redirects *.html and
// DROPS the query string -> read the fragment (#...) too, which survives.
var raw = location.search ? location.search.slice( 1 ) : location.hash.slice( 1 );
var p = new URLSearchParams( raw );
function gi( name )
{
var v = p.get( name );
if( !v )
return -2147483648; // INT_MIN sentinel
var x = parseInt( v, 10 );
return isNaN( x ) ? -2147483648 : x;
}
var b = $0 >> 2;
HEAP32[b + 0] = gi( 'm' );
HEAP32[b + 1] = gi( 'park' );
HEAP32[b + 2] = gi( 'work' );
HEAP32[b + 3] = gi( 'blocks' );
HEAP32[b + 4] = gi( 'iters' );
HEAP32[b + 5] = gi( 'passes' );
console.log( '[RTPOOL] location.search=' + location.search );
},
out );
#endif
}
static long elapsedMs( clock_t_::time_point t0 )
{
return (long) std::chrono::duration_cast<std::chrono::milliseconds>(
clock_t_::now() - t0 ).count();
}
// ----------------------------------------------------------------------------
// the shared "work" — mirrors the raytracer: workers cooperatively drain a
// global atomic block counter; each block is pure CPU math (no JS/DOM/wx — the
// reason the real raytracer workers are safe to run off the main thread).
// ----------------------------------------------------------------------------
static std::atomic<size_t> g_nextBlock{ 0 };
static std::atomic<size_t> g_threadsFinished{ 0 };
static std::atomic<int> g_workersRan{ 0 }; // distinct workers that ran >=1 block
static std::atomic<double> g_sink{ 0.0 }; // keep the optimizer honest
static int g_numBlocks = 64;
static long g_iters = 4000000;
static double computeBlock( int b, long iters )
{
double acc = 0.0;
for( long i = 0; i < iters; ++i )
acc += std::sin( (double) ( b * 131 + i ) * 1e-6 ) * std::cos( (double) i * 1e-7 );
return acc;
}
static void workerBody()
{
bool counted = false;
for( ;; )
{
size_t b = g_nextBlock.fetch_add( 1 );
if( b >= (size_t) g_numBlocks )
break;
if( !counted )
{
g_workersRan.fetch_add( 1 );
counted = true;
}
double r = computeBlock( (int) b, g_iters );
g_sink.store( g_sink.load( std::memory_order_relaxed ) + r, std::memory_order_relaxed );
}
g_threadsFinished.fetch_add( 1 );
}
static void resetPass()
{
g_nextBlock.store( 0 );
g_threadsFinished.store( 0 );
g_workersRan.store( 0 );
}
// ----------------------------------------------------------------------------
// "parked" pool — block N workers on a condition variable (zero CPU) to occupy
// the pre-warmed pool slots, exactly like KiCad's singleton thread pool.
// ----------------------------------------------------------------------------
static std::mutex g_parkMtx;
static std::condition_variable g_parkCv;
static bool g_parkRelease = false;
static void spawnParkedPool( int p )
{
for( int i = 0; i < p; ++i )
{
std::thread(
[]
{
std::unique_lock<std::mutex> lk( g_parkMtx );
g_parkCv.wait( lk, [] { return g_parkRelease; } );
} )
.detach();
}
}
static void releaseParkedPool()
{
{
std::lock_guard<std::mutex> lk( g_parkMtx );
g_parkRelease = true;
}
g_parkCv.notify_all();
}
// ----------------------------------------------------------------------------
// variant A: detached threads + std::this_thread::sleep_for busy-wait.
// Returns true if a deadlock was detected (workers never finished within 12s).
// ----------------------------------------------------------------------------
static bool waitBusy( int n )
{
auto start = clock_t_::now();
while( g_threadsFinished.load() < (size_t) n )
{
std::this_thread::sleep_for( std::chrono::milliseconds( 10 ) );
if( std::chrono::duration_cast<std::chrono::seconds>( clock_t_::now() - start ).count() >= 12 )
return true; // deadlock: the event loop never ran, workers never started
}
return false;
}
// ----------------------------------------------------------------------------
// variant B1: detached threads + emscripten_sleep yield-poll.
// ----------------------------------------------------------------------------
static void waitYield( int n )
{
while( g_threadsFinished.load() < (size_t) n )
{
#ifdef __EMSCRIPTEN__
emscripten_sleep( 10 ); // Asyncify: hand control back to the event loop
#else
std::this_thread::sleep_for( std::chrono::milliseconds( 10 ) );
#endif
}
}
// ----------------------------------------------------------------------------
// variant B2: a persistent pre-warmed worker pool. Workers are created once and
// block on a condition variable between passes (blocking is fine on workers —
// only the MAIN thread can't block). Each pass: main bumps the generation,
// notifies, then emscripten_sleep-polls the done counter.
// ----------------------------------------------------------------------------
struct PersistentPool
{
int n = 0;
std::vector<std::thread> ts;
std::mutex mtx;
std::condition_variable cv;
unsigned long generation = 0;
bool stop = false;
std::atomic<size_t> finished{ 0 };
std::atomic<size_t> started{ 0 }; // workers that have come alive
void start( int n_ )
{
n = n_;
for( int i = 0; i < n; ++i )
ts.emplace_back( [this] { loop(); } );
}
bool ready() const { return started.load() == (size_t) n; }
void loop()
{
started.fetch_add( 1 );
unsigned long seen = 0;
for( ;; )
{
std::unique_lock<std::mutex> lk( mtx );
cv.wait( lk, [&] { return stop || generation != seen; } );
if( stop )
return;
seen = generation;
lk.unlock();
workerBody(); // drains g_nextBlock cooperatively with the other workers
finished.fetch_add( 1 );
}
}
void runPass() // called on the MAIN thread
{
finished.store( 0 );
resetPass();
{
std::lock_guard<std::mutex> lk( mtx );
++generation;
}
cv.notify_all();
while( finished.load() < (size_t) n )
{
#ifdef __EMSCRIPTEN__
emscripten_sleep( 5 );
#else
std::this_thread::sleep_for( std::chrono::milliseconds( 5 ) );
#endif
}
}
// Like runPass(), but waits with a BUSY sleep_for (NOT emscripten_sleep). This is
// the mechanism the real raytracer uses: the workers are already alive, so a
// main-thread busy-wait still completes (they run on their own cores), and there
// is NO emscripten_sleep — so it can't hit the Asyncify-nesting abort that the
// real 3D viewer triggers under its modal/event-pump. REQUIRES ready()==true.
void runPassBusyWait() // called on the MAIN thread
{
finished.store( 0 );
resetPass();
{
std::lock_guard<std::mutex> lk( mtx );
++generation;
}
cv.notify_all();
while( finished.load() < (size_t) n )
std::this_thread::sleep_for( std::chrono::microseconds( 200 ) );
}
};
static PersistentPool g_pool;
// ----------------------------------------------------------------------------
// app
// ----------------------------------------------------------------------------
class RtFrame : public wxFrame
{
public:
RtFrame() : wxFrame( nullptr, wxID_ANY, "Raytrace Threads Repro",
wxDefaultPosition, wxSize( 420, 160 ) )
{
wxPanel* panel = new wxPanel( this );
wxBoxSizer* sizer = new wxBoxSizer( wxVERTICAL );
sizer->Add( new wxStaticText( panel, wxID_ANY,
"Raytrace threading repro — see the console." ),
0, wxALL, 16 );
panel->SetSizer( sizer );
}
};
class RtApp : public wxApp
{
public:
bool OnInit() override
{
const int hwc = std::max( 1u, std::thread::hardware_concurrency() );
int raw[6];
readUrlParams( raw );
auto pick = [&]( int i, int dflt ) { return raw[i] == INT_MIN ? dflt : raw[i]; };
const int mode = pick( 0, 0 );
const int park = pick( 1, hwc );
const int work = pick( 2, hwc );
g_numBlocks = pick( 3, 64 );
g_iters = (long) pick( 4, 4000000 );
const int passes = pick( 5, 1 );
rtlog( "[RTPOOL] START m=%d park=%d work=%d blocks=%d iters=%ld passes=%d hwc=%d",
mode, park, work, g_numBlocks, g_iters, passes, hwc );
// exhaust the pre-warmed pool (skip for the pure-serial baseline)
if( mode != 3 && park > 0 )
{
spawnParkedPool( park );
rtlog( "[RTPOOL] parked %d workers (pool slots now exhausted)", park );
}
auto t0 = clock_t_::now();
bool deadlocked = false;
for( int pass = 0; pass < passes && !deadlocked; ++pass )
{
auto p0 = clock_t_::now();
switch( mode )
{
case 3: // C — serial on the calling (main) thread
resetPass();
workerBody();
break;
case 0: // A — detached threads + busy-wait (expected deadlock)
resetPass();
for( int i = 0; i < work; ++i )
std::thread( workerBody ).detach();
deadlocked = waitBusy( work );
break;
case 1: // B1 — detached threads + emscripten_sleep yield
resetPass();
for( int i = 0; i < work; ++i )
std::thread( workerBody ).detach();
waitYield( work );
break;
case 2: // B2 — persistent pool + emscripten_sleep yield
if( pass == 0 )
g_pool.start( work );
g_pool.runPass();
break;
case 4: // B1 but with STACK-LOCAL atomics — mirrors the real raytracer
// EXACTLY (threadsFinished/nextBlock are locals shared by ref with
// the workers). Proves Asyncify's emscripten_sleep unwind/rewind
// doesn't clobber concurrent worker writes to C-stack locals.
{
std::atomic<size_t> lNext( 0 );
std::atomic<size_t> lFinished( 0 );
std::atomic<int> lRan( 0 );
auto localWorker = [&]()
{
bool counted = false;
for( ;; )
{
size_t b = lNext.fetch_add( 1 );
if( b >= (size_t) g_numBlocks )
break;
if( !counted )
{
lRan.fetch_add( 1 );
counted = true;
}
g_sink.store( g_sink.load( std::memory_order_relaxed )
+ computeBlock( (int) b, g_iters ),
std::memory_order_relaxed );
}
lFinished.fetch_add( 1 );
};
for( int i = 0; i < work; ++i )
std::thread( localWorker ).detach();
while( lFinished.load() < (size_t) work )
#ifdef __EMSCRIPTEN__
emscripten_sleep( 2 );
#else
std::this_thread::sleep_for( std::chrono::milliseconds( 10 ) );
#endif
g_workersRan.store( lRan.load() );
break;
}
case 5: // B3 — persistent pool + sleep_for BUSY-WAIT. This is the REAL
// raytracer's mechanism: workers are pre-alive, so a main-thread
// busy-wait still completes (no event-loop dependency), and there is
// NO emscripten_sleep, so it can't hit the Asyncify-nesting abort the
// 3D viewer triggers under its modal/event-pump.
if( pass == 0 )
{
g_pool.start( work );
// Warm up: the workers must actually be created (needs the event loop).
// In the real app this happens across paint frames; here we yield once.
while( !g_pool.ready() )
#ifdef __EMSCRIPTEN__
emscripten_sleep( 5 );
#else
std::this_thread::sleep_for( std::chrono::milliseconds( 5 ) );
#endif
}
g_pool.runPassBusyWait();
break;
}
long passMs = elapsedMs( p0 );
if( deadlocked )
rtlog( "[RTPOOL] DEADLOCK pass=%d workersRan=%d threadsFinished=%d "
"(workers never ran within 12s)",
pass, g_workersRan.load(), (int) g_threadsFinished.load() );
else
rtlog( "[RTPOOL] PASS done pass=%d workersRan=%d passMs=%ld",
pass, g_workersRan.load(), passMs );
}
long totalMs = elapsedMs( t0 );
if( deadlocked )
rtlog( "[RTPOOL] FAIL deadlocked mode=%d totalMs=%ld", mode, totalMs );
else
rtlog( "[RTPOOL] SUCCESS mode=%d workersRan=%d totalMs=%ld sink=%.3f",
mode, g_workersRan.load(), totalMs, g_sink.load() );
releaseParkedPool();
( new RtFrame() )->Show();
return true;
}
};
wxIMPLEMENT_APP( RtApp );