Build the prepared shaders into the driver's cache at an update only when a sample says the driver lost them.

After an update that keeps the prepared shaders, --warm-shaders built all 42,000 of them into the graphics driver's
cache again: 10 to 20 s at every update on a player's PC while the driver gave every one back at once, and the full
build when it no longer had them (380 s there once, after a change of what they are built with). It now builds twelve
fragment libraries spread over the list first, into the driver's cache alone, as the game's check at every start
does, and builds the whole list only when fewer than three quarters of them came from that cache.

On copies of an install's caches (RTX 5090): with the driver's cache partly cold after a relink, 5 of 12 came from it
and the whole list was built (313 s); the run after it found 12 of 12 and built nothing more, the step taking 3 s.
lib_warm = force builds the whole list whatever the sample says.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
chrg6
2026-11-13 03:40:21 -05:00
parent a8332948f2
commit 13e6563236
2 changed files with 50 additions and 7 deletions

View File

@@ -8026,8 +8026,35 @@ static void lb_warm_worker(void *arg) {
}
}
// Preparation: every variant of <dir>/shader-prepared.bin on every core at normal priority.
static void lb_warm_prepared(const char *dir) {
// --warm-shaders, before the whole list: LW_SAMPLE prepared fragment libraries spread over it, built into the driver's
// own cache alone (no_cache, as the library warm in play builds) and timed, as the game's check at every start does
// (lw_check). 1 when at least LW_WARM_SHARE of them came from the driver's cache. *fast and *got: the counts.
static int prep_sample_warm(int *fast, int *got) {
double ms[LW_SAMPLE];
*fast = *got = 0;
size_t step = LB.n / LW_SAMPLE ? LB.n / LW_SAMPLE : 1;
int was = no_cache;
no_cache = 1;
for (size_t i = step / 2; i < LB.n && *got < LW_SAMPLE; i += step)
for (size_t j = i; j < i + step && j < LB.n; j++) { // the next fragment variant from here
if (LB.e[j].stage != GFX_STAGE_FS) continue;
journal_mark(0, &LB.e[j]); // a driver that dies on it names it, as the full pass would
double t = lw_build(&LB.e[j]);
journal_mark(0, NULL);
if (t >= 0) ms[(*got)++] = t;
break;
}
no_cache = was;
for (int i = 0; i < *got; i++) *fast += ms[i] <= LW_COLD_MS;
return *got >= LW_SAMPLE / 2 && lw_warm_sample(*fast, *got);
}
// Preparation: every variant of <dir>/shader-prepared.bin on every core at normal priority. sample: --warm-shaders
// after an update, which first asks the driver's cache (prep_sample_warm) and builds nothing more when it has them.
// It took 10 to 20 s on a player's PC at every update, all of it 42,000 libraries the driver gave back at once, and
// it still takes the minutes of a full build when the driver no longer has them (a driver update, or a change of what
// they are built with: 380 s on that PC once). NX2NAVI_LIB_WARM=force builds the whole list whatever the sample says.
static void lb_warm_prepared(const char *dir, int sample) {
char path[PLAT_PATH_MAX * 2 + 32];
snprintf(path, sizeof path, "%s/shader-prepared.bin", dir);
size_t n = 0;
@@ -8045,6 +8072,20 @@ static void lb_warm_prepared(const char *dir) {
LB.n = n, LB.idle = 0, LB.t0 = now_ms();
static const uint32_t no_pr[16], no_fs[4];
lb_state(no_pr, no_fs);
int fast = 0, got = 0;
if (sample && P.gpl && strcmp(cfg_value("NX2NAVI_LIB_WARM"), "force") && prep_sample_warm(&fast, &got)) {
if (SK.journal) plat_file_close(SK.journal), SK.journal = NULL, plat_file_delete(jp);
printf("prepare-shaders: progress build %zu %zu\n", n, n);
printf("prepare-shaders: the graphics driver's cache still has the prepared shaders (%d of %d sampled came from "
"it in under %.0f ms each): the %zu variants are not built again (%.0f ms)\n", fast, got, LW_COLD_MS, n,
now_ms() - LB.t0);
fflush(stdout);
free(LB.e), LB.e = NULL, LB.n = 0, LB.next = 0, LB.done = 0, LB.skipped = 0;
return;
}
if (sample && got)
printf("prepare-shaders: the graphics driver's cache does not have the prepared shaders (%d of %d sampled came "
"from it): all %zu variants are built into it again\n", fast, got, n);
int cpus = (int)plat_cpu_usable(), k = 0;
PlatThread *t[JR_SLOTS];
for (int i = 0; i < cpus && i < JR_SLOTS; i++)
@@ -8281,7 +8322,7 @@ int vk_pipeline_prepare_shaders(void) {
// The listed pipelines before the stage libraries: the pipeline cache is saved with them in it and without the
// tens of thousands of libraries the next step builds, which would take the file past its size limit.
if (!rc) seed_pipelines(dir);
if (!rc) lb_warm_prepared(dir);
if (!rc) lb_warm_prepared(dir, 0);
return rc;
}
@@ -8304,7 +8345,7 @@ int vk_pipeline_warm_shaders(void) {
return 1;
}
seed_pipelines(dir); // before the libraries, as in vk_pipeline_prepare_shaders
lb_warm_prepared(dir);
lb_warm_prepared(dir, 1);
return 0;
}