Merge main into the 0.9.1 release: the launcher's animations stopped behind another window, the uniform snapshot's copy split over the replay workers, and the replay workers' default of three with its measurements.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
This commit is contained in:
chrg6
2026-11-12 07:29:41 -05:00
5 changed files with 103 additions and 12 deletions

View File

@@ -1761,6 +1761,9 @@ static void op_one(VkExec *ex, const op_hdr *h) {
}
// ---- replay workers (NX2NAVI_REPLAY_WORKERS, default 3; 0: the replay records every op on the submitting thread)
// Default 3: on an RX 7900 XT with a Ryzen 7 5800X3D, Hyrule Field standing went from 41.2 to 50.4 fps (2026-11-12, one
// batch), with the same pictures there and in Kokiri Forest, Kakariko, Goron City, a graveyard house and Gerudo
// Valley; 5 or 7 workers measured the same as 3.
// A submission's draws come in runs between ops that have to stay in the replay's order on its own thread (dispatches,
// copies, texture and buffer clears, events, the game's barriers, calls). Hyrule Field, standing (NX2NAVI_REPLAY_RUNS,
// 2026-11-11): about 3,500 draws a frame in 82 runs, 96% of them in runs of 16 draws or more and 67% in runs of 256 or
@@ -1879,12 +1882,66 @@ static int ser_kind(uint32_t op) {
static Chunk *chunk_take(void);
static void chunk_record(Chunk *c, int pool);
// ---- side work (vk_exec_par): the replay thread hands the workers a loop over n items while they have no chunk to
// record, at a submission's start (vk_usnap.c copy_regions). One job at a time, in a static slot: seq is odd while it
// runs. A worker counts itself in (PJ.in) and checks seq again before it reads the job, and the caller, once every
// slice is done, makes seq even and waits until no worker is in, so none reads a job that has ended.
static struct {
void (*fn)(void *, int, int);
void *arg;
int n, grain;
volatile int32_t next, left, in;
volatile uint32_t seq;
} PJ;
static PlatMutex pj_lock = PLAT_MUTEX_INIT;
static void par_slices(void) {
for (;;) {
int lo = __atomic_fetch_add(&PJ.next, PJ.grain, __ATOMIC_ACQ_REL);
if (lo >= PJ.n) return;
int hi = lo + PJ.grain < PJ.n ? lo + PJ.grain : PJ.n;
PJ.fn(PJ.arg, lo, hi);
__atomic_sub_fetch(&PJ.left, hi - lo, __ATOMIC_ACQ_REL);
}
}
static int par_pending(void) { return __atomic_load_n(&PJ.seq, __ATOMIC_ACQUIRE) & 1; }
static void par_help(void) {
uint32_t s = __atomic_load_n(&PJ.seq, __ATOMIC_ACQUIRE);
if (!(s & 1)) return;
__atomic_add_fetch(&PJ.in, 1, __ATOMIC_SEQ_CST);
if (__atomic_load_n(&PJ.seq, __ATOMIC_SEQ_CST) == s) par_slices();
__atomic_sub_fetch(&PJ.in, 1, __ATOMIC_SEQ_CST);
}
void vk_exec_par(void (*fn)(void *, int, int), void *arg, int n, int grain) {
if (grain < 1) grain = 1;
if (RW.n <= 0 || n <= grain || Q.drained != Q.out) {
if (n > 0) fn(arg, 0, n);
return;
}
plat_mutex_lock(&pj_lock);
PJ.fn = fn, PJ.arg = arg, PJ.n = n, PJ.grain = grain;
__atomic_store_n(&PJ.next, 0, __ATOMIC_RELAXED);
__atomic_store_n(&PJ.left, n, __ATOMIC_RELAXED);
__atomic_add_fetch(&PJ.seq, 1, __ATOMIC_SEQ_CST); // odd: running
for (int k = 0; k < RW.n; k++)
if (__atomic_load_n(&RW.w[k]->sleeping, __ATOMIC_SEQ_CST)) plat_event_set(&RW.w[k]->wake);
par_slices();
while (__atomic_load_n(&PJ.left, __ATOMIC_ACQUIRE) > 0) plat_pause();
__atomic_add_fetch(&PJ.seq, 1, __ATOMIC_SEQ_CST); // even: ended
while (__atomic_load_n(&PJ.in, __ATOMIC_SEQ_CST)) plat_pause();
plat_mutex_unlock(&pj_lock);
}
static void rw_main(void *arg) {
RwWorker *w = arg;
plat_thread_priority(PLAT_PRIO_HIGHEST); // as the recording thread (rec_claim)
perf_lock_timing_here(); // NX2NAVI_LOCK_TIME, as on the recording thread
vk_exec_draw_ctx_use(w->dc);
for (;;) {
par_help();
Chunk *c = chunk_take();
if (c) {
chunk_record(c, w->idx);
@@ -1892,13 +1949,16 @@ static void rw_main(void *arg) {
}
// spins about 100 us after a chunk (the next group of the frame is usually closer), then sleeps
uint64_t t0 = kern_now_ns();
for (unsigned i = 0; __atomic_load_n(&Q.take, __ATOMIC_SEQ_CST) >= __atomic_load_n(&Q.put, __ATOMIC_SEQ_CST); i++) {
for (unsigned i = 0; __atomic_load_n(&Q.take, __ATOMIC_SEQ_CST) >= __atomic_load_n(&Q.put, __ATOMIC_SEQ_CST);
i++) {
if (par_pending()) break;
if ((i & 63) || kern_now_ns() - t0 < 100000) {
plat_pause();
continue;
}
__atomic_store_n(&w->sleeping, 1, __ATOMIC_SEQ_CST);
if (__atomic_load_n(&Q.take, __ATOMIC_SEQ_CST) >= __atomic_load_n(&Q.put, __ATOMIC_SEQ_CST))
if (__atomic_load_n(&Q.take, __ATOMIC_SEQ_CST) >= __atomic_load_n(&Q.put, __ATOMIC_SEQ_CST) &&
!par_pending())
plat_event_wait_ns(&w->wake, 2000000);
__atomic_store_n(&w->sleeping, 0, __ATOMIC_RELAXED);
t0 = kern_now_ns();