Leave the NDSR upscaler's slow, unused compute shaders out of the install's shader preparation.

The translator turns the larger programs of agl_ndsr.pck and agl_ndsr_vr.pck into 3-12.6 MB of SPIR-V, which the GPU driver took 34 s to 44 minutes each to build: 19,400 of the 21,750 s every compute variant took together, most of the install's hour of driver builds. Play met none of them that was slow to build, so the default variants of those programs above 16 KB of code are left out; agl_ndsr_vr_light.pck, whose programs play meets, stays whole. --all-shaders prepares them too.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

tools/prepare_shaders.py

@@ -63,6 +63,17 @@ DEFAULT = {'tex': {str(slot): '2d' for slot in range(32)}}   # answers outside a
 CONTROL = 0x880
 ANS_REC = 4 + 32 + 32 + 147 + 1
 DEFAULT_FLAG = 0x80   # pad byte: a default variant, its answers refined from its first translation (gfx_shader_guess.c)
+# Default variants left out unless --all-shaders: the larger compute programs of two of Ocarina's three NDSR upscaler
+# packages. The translator turns their 20-47 KB of code into 3-12.6 MB of SPIR-V, which the GPU driver took 34 s to
+# 44 min each to build (RTX 5090, driver 621.01): 35 programs, the 24 that translate took 19,400 of the 21,750 s
+# every compute variant together took. Play sessions met none of them that was slow to build (one 20.8 KB program built in 0.2 s), while all five
+# agl_ndsr_vr_light.pck programs play met took 77-153 s each, so that package stays whole. A program play does meet
+# here is translated and built at its first use, as before the NDSR packages were read at all.
+SLOW_UNUSED = (('/agl_ndsr.pck', 16 << 10), ('/agl_ndsr_vr.pck', 16 << 10))   # (source suffix, code bytes above which)
+
+
+def slow_unused(s):
+    return any(s['source'].endswith(suffix) and len(s['code']) > limit for suffix, limit in SLOW_UNUSED)
 JOURNAL = 'shader-prepare-journal.bin'   # vk_pipeline.c skip_init: a 64-byte slot per worker, input id first
 RETRIES = 3
 
@@ -179,6 +190,8 @@ def main():
     ap.add_argument('--list', help='a shader_list.tsv instead of tools/data/shader_list.tsv or refs/MacNativeGames')
     ap.add_argument('--no-run', action='store_true', help='write the seed files only')
     ap.add_argument('--no-list', action='store_true', help='default variants only, without a shared shader list')
+    ap.add_argument('--all-shaders', action='store_true',
+                    help='also the default variants SLOW_UNUSED leaves out (an hour more of GPU driver builds)')
     ap.add_argument('--exe', default='nx2navi.exe' if os.name == 'nt' else 'nx2navi',
                     help='the game executable in the build directory that runs --prepare-shaders')
     a = ap.parse_args()
@@ -196,7 +209,8 @@ def main():
         shaders, _ = inventory(a.romfs or extracted / 'romfs', a.exefs or extracted / 'exefs')
     t1 = time.monotonic()
     progs, recs, seen, skipped = {}, [], set(), {'not in this version': 0, 'stage': 0}
-    variants += [(key, DEFAULT) for key in sorted(shaders)]
+    left_out = 0 if a.all_shaders else sum(1 for s in shaders.values() if slow_unused(s))
+    variants += [(key, DEFAULT) for key in sorted(shaders) if a.all_shaders or not slow_unused(shaders[key])]
     for key, deps in variants:
         s = shaders.get(key)
         if not s:
@@ -233,7 +247,8 @@ def main():
             f.write(rec)
     by = [sum(1 for p, _ in recs if p == i) for i in range(3)]
     print(f'{listed} listed variants and {len(shaders)} default ones: {len(recs)} of {len(progs)} shaders in {a.version} '
-          f'(priority t {by[0]}, warm {by[1]}, rest {by[2]}), skipped {skipped}; inventory {t1 - t0:.1f} s, '
+          f'(priority t {by[0]}, warm {by[1]}, rest {by[2]}), skipped {skipped}, {left_out} slow unused default variants '
+          f'left out (SLOW_UNUSED); inventory {t1 - t0:.1f} s, '
           f'seed written in {time.monotonic() - t0:.1f} s', flush=True)
     if a.no_run:
         return 0