diff --git a/nx2navi.ini.example b/nx2navi.ini.example index 3e69e8b2..ace664f9 100644 --- a/nx2navi.ini.example +++ b/nx2navi.ini.example @@ -364,6 +364,9 @@ # With astc = gpu: 1 keeps the decoder's scratch buffer from one upload batch to the next instead of making it again for # each, to shorten the hitches when an area loads (0 = one per batch). # astc_scratch_keep = 1 +# With astc_scratch_keep = 1: 1 gives back a scratch grown for a large texture 5 s after the last one needed it (up to +# 192 MiB of video memory after an area load); 0 keeps it until the next ASTC upload, as in 0.10.1. +# astc_scratch_trim = 1 # With astc = gpu: 1 stages a texture's ASTC blocks in the shared staging ranges instead of a buffer of their own # (0 = a buffer each). # astc_arena = 1 diff --git a/runtime/cfg.c b/runtime/cfg.c index 04406032..0e709686 100644 --- a/runtime/cfg.c +++ b/runtime/cfg.c @@ -61,7 +61,7 @@ static const char *const known[] = { "SKIN_MOTION", // vk_skin_motion.inc "MVOBJ_ALL", "MVOBJ_BANKS", "MVOBJ_EQUAL", "MVOBJ_EXACT", "MVOBJ_EXEC", "MVOBJ_NOC4", "MVOBJ_ONLY", "MVOBJ_TIMES", "MVPREV_CHECK", "MVPREV_GROW", "MVPREV_SKIP", "MVPREV_TRUST", "MV_CHECK", "MV_SURVEY", - "PIPE_SEED_BOOT", "ONE_PER_CORE", "OPERATION_MODE", "OUTPUT_RES", "PAD_LAYOUT", "PACE_HZ", "PASS_CRUMBS", "PATCH_OUTPUT", "PATCH_RES_FIXED", "PATCH_SCENE", "PATCH_WIDE_ASPECT", "PATCH_WIDE_HUD", "PATCH_WIDE_MOVIE", "PATCH_WIDE_OUT", "PATCH_FREECAM", "PATCH_OCCLUSION_OFF", "PATCH_WIDE_SAVE", "PATCH_WIDE_SCENE", "PCSS", "PCSS_K", "PERF_OVERLAY", "PERF_SLOW_MS", "PIPELINE_CACHE", "PIPELINE_CACHE_MAX_MB", "PIPELINE_CACHE_PLAY", "PIPELINE_CACHE_PLAY_MB", "PIPE_BINARIES", "PIPE_BINARIES_MAX_MB", "PIPE_FEEDBACK", "PIPE_MEMO", "PIPE_PREBUILD", "PIPE_SEED", "PIPE_SKIP", "PIPE_SYNC", "PIPE_WAIT", "POWER_FULL", "TEX_SAME", "TEX_SAME_BIG", "TEX_FLAT3D", "TEX_KEEP_TAKEN", "TEX_HARD_ORDER", "TEX_STALE_KEPT", "PRESENT_HELD", "HELD_SYNC", + "PIPE_SEED_BOOT", "ONE_PER_CORE", "OPERATION_MODE", "OUTPUT_RES", "PAD_LAYOUT", "PACE_HZ", "PASS_CRUMBS", "PATCH_OUTPUT", "PATCH_RES_FIXED", "PATCH_SCENE", "PATCH_WIDE_ASPECT", "PATCH_WIDE_HUD", "PATCH_WIDE_MOVIE", "PATCH_WIDE_OUT", "PATCH_FREECAM", "PATCH_OCCLUSION_OFF", "PATCH_WIDE_SAVE", "PATCH_WIDE_SCENE", "PCSS", "PCSS_K", "PERF_OVERLAY", "PERF_SLOW_MS", "PIPELINE_CACHE", "PIPELINE_CACHE_MAX_MB", "PIPELINE_CACHE_PLAY", "PIPELINE_CACHE_PLAY_MB", "PIPE_BINARIES", "PIPE_BINARIES_MAX_MB", "PIPE_FEEDBACK", "PIPE_MEMO", "PIPE_PREBUILD", "PIPE_SEED", "PIPE_SKIP", "PIPE_SYNC", "PIPE_WAIT", "POWER_FULL", "TEX_SAME", "TEX_SAME_BIG", "TEX_FLAT3D", "TEX_KEEP_TAKEN", "TEX_HARD_ORDER", "TEX_STALE_KEPT", "ASTC_SCRATCH_TRIM", "PRESENT_HELD", "HELD_SYNC", "PREPARE_PIPELINES", "PREWARM_FULL", "PREWARM_FULL_S", "PREWARM_WAIT_MS", "PREWARM_XLATE", "PROBE_MOVIE", "PROBE_MOVIE_SHOT", "PROFILE", "PROFILE_CHAINS", "PROFILE_DIR", "PROFILE_FRAMES", "PROFILE_REPLAY", "PROFILE_SLOW", "POOL_FAST", "POOL_GROUP", "POOL_GROUP_CHECK", "POOL_PRESNAP", "POOL_PRESNAP_RUNS", "SUBMIT_FINISH_GPU", "SVC_WAITS", "SNAP_WIDEN", "SUBMIT_FINISH_RESULTS", "SUBMIT_FINISH_KINDS", "FOLIAGE", "MIRROR_STAGE", "FLUSH_SKIP_CLEAN", "MIRROR_HOT", "MESH_EMULATION", "MESH_EMU_MB", "MESH_EMU_TASKS", "MESH_EMU_BATCH", "GUEST_OPT", "LOG_TEX_NEW", "DUMP_KEY", "VEC_INDEX_FIX", "PASS_TIMES", diff --git a/runtime/vk.h b/runtime/vk.h index 9f3a041b..3cf5edb9 100644 --- a/runtime/vk.h +++ b/runtime/vk.h @@ -598,6 +598,7 @@ void vk_texture_split_counts(uint64_t *copies, uint64_t *fills); VkDeviceSize vk_astc_region_bytes(const VkAstcRegion *r, uint32_t bw, uint32_t bh); // source bytes the region reads void vk_astc_stats(uint64_t *dispatches, uint64_t *texels); // since start uint64_t vk_astc_scratch_bytes(void); // the decoder's shared scratch now +void vk_astc_scratch_trim(void); // each present: a grown scratch idle for 5 s goes // NX2NAVI_ASTC_TIME=1: GPU timestamps around each vk_astc_record's dispatches (logged per footprint and the largest one) void vk_astc_time_enable(int on); int vk_astc_time_get(uint32_t bw, uint32_t bh, uint64_t *texels, uint64_t *calls, double *us); // totals so far diff --git a/runtime/vk_astc.c b/runtime/vk_astc.c index 693cb795..58b7e086 100644 --- a/runtime/vk_astc.c +++ b/runtime/vk_astc.c @@ -487,7 +487,7 @@ static void mem_barrier(VkCommandBuffer cb, VkPipelineStageFlags2 src, VkAccessF // only after that batch completed (vk_memory.c: completed batches go to the spare list), so the same handle means the // open batch, ordered by a barrier, or a finished one. g_lock held. static struct { VkBuffer buf; VmaAllocation alloc; VkDeviceSize size; VkCommandBuffer cb; VkDevice dev; uint64_t big_ns; } S; -static uint64_t n_scratch_new, n_scratch_reuse; +static uint64_t n_scratch_new, n_scratch_reuse, n_scratch_trim; // NX2NAVI_ASTC_SCRATCH_KEEP=1 (the default): the scratch outlives its upload batch. Every // batch goes to gfx_queue, so the reuse barrier at the start of the next batch's first chunk (vk_astc_record) orders it @@ -560,6 +560,27 @@ static VkBuffer scratch_get(VkCtx *c, VkCommandBuffer cb, VkDeviceSize need, int return out; } +// Every present (vk_texture.c vk_texstat_frame): a scratch grown past SCRATCH_MIN goes once no texture needed more for +// SCRATCH_SHRINK_NS, without waiting for the next ASTC upload, which is when scratch_get makes it small again: in a +// session of 2026-11-13 the decoder held 224 MiB for the 70 s after an area load and from a later load to the session's +// end. Nothing is made here: the next request makes the usual scratch, as after scratch_get's own shrink. The buffer +// is freed once the upload batches recorded so far completed (vk_defer_buffer), as there. Only tried: scratch_get holds +// g_lock while VMA makes a buffer, up to 19 ms for a new block, and the present does not wait for that. +// NX2NAVI_ASTC_SCRATCH_TRIM=0: the grown scratch stays until the next upload, as in 0.10.1. +void vk_astc_scratch_trim(void) { + static int on = -1; + if (on < 0) on = rt_knob("ASTC_SCRATCH_TRIM", "1") == 1; + if (!on || !scratch_keep()) return; + if (__atomic_load_n(&S.size, __ATOMIC_RELAXED) <= SCRATCH_MIN || !plat_mutex_trylock(&g_lock)) return; + VkCtx *c = vk_ctx(); + if (S.buf && c && S.dev == c->device && S.size > SCRATCH_MIN && kern_now_ns() - S.big_ns > SCRATCH_SHRINK_NS) { + vk_defer_buffer(S.buf, S.alloc); + S.buf = VK_NULL_HANDLE, S.alloc = NULL, S.size = 0, S.cb = VK_NULL_HANDLE; + n_scratch_trim++; + } + plat_mutex_unlock(&g_lock); +} + // The shared scratch's size, for the log's VRAM breakdown (scratches still owned by an upload batch not counted). uint64_t vk_astc_scratch_bytes(void) { plat_mutex_lock(&g_lock); @@ -762,9 +783,9 @@ int vk_astc_record(VkCommandBuffer cb, VkBuffer src, VkDeviceSize src_size, cons } if ((n_dispatch + (uint64_t)n) / 1024 != n_dispatch / 1024 || !n_dispatch) { rt_log("vk: ASTC: %llu regions, %.1f M texels decoded on the GPU so far; shared scratch %llu MiB, made %llu " - "times, reused %llu", (unsigned long long)(n_dispatch + (uint64_t)n), (double)n_texels / 1e6, - (unsigned long long)(S.size >> 20), (unsigned long long)n_scratch_new, - (unsigned long long)n_scratch_reuse); + "times, reused %llu, given back idle %llu", (unsigned long long)(n_dispatch + (uint64_t)n), + (double)n_texels / 1e6, (unsigned long long)(S.size >> 20), (unsigned long long)n_scratch_new, + (unsigned long long)n_scratch_reuse, (unsigned long long)n_scratch_trim); vk_astc_time_summary(); } n_dispatch += (uint64_t)n; diff --git a/runtime/vk_texture.c b/runtime/vk_texture.c index 9a73eacb..f5e846d2 100644 --- a/runtime/vk_texture.c +++ b/runtime/vk_texture.c @@ -3429,6 +3429,7 @@ static void compact_frame_time(uint64_t d); void vk_texstat_frame(uint64_t frame, uint64_t d) { if (compact_on()) compact_frame_time(d); + vk_astc_scratch_trim(); if (ts_log_ms < 0) { ts_log_ms = cfg_int("NX2NAVI_TEXUP_LOG_MS"); if (ts_log_ms < 0) ts_log_ms = 0;