Triforce v0.3.0

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
This commit is contained in:
ruckus7382sfgsgdf
2026-11-11 20:19:23 -06:00
parent 035977a061
commit a939e9d0d1
12 changed files with 1998 additions and 107 deletions

View File

@@ -1190,11 +1190,46 @@ void state_record(const Program& p, const NvnState& s, const Targets& at, VkPrim
if (!state_file) state_open_locked();
if (state_file && state_seen.insert(h).second) fwrite(&r, sizeof r, 1, state_file), fflush(state_file);
}
/// (perf push 11-11) Which shipped list this folder's notes were last made from, in triforce_states_applied.bin: the list's
/// contents and the translator, hashed. The lists were gone through only inside a prepare (a first start, a new translator, a
/// new seed): a folder prepared already that got a longer shipped list (an update with the same translator, a list copied in)
/// never noted its pipelines, and built each in play when first drawn (a lab folder with a 12067-pipeline list: 2687 built in
/// play, the list not read once). 0: there is no shipped list
uint64_t state_shipped_id() {
FILE* f = fopen(state_list_path("triforce_pipeline_states_shipped.bin").c_str(), "rb");
if (!f) return 0;
uint64_t h = 0xCBF29CE484222325ull ^ program_cache_stamp(), total = 0;
std::vector<uint8_t> buf(1 << 16);
for (size_t n; (n = fread(buf.data(), 1, buf.size(), f)) > 0; total += n)
for (size_t i = 0; i < n; i++) h = (h ^ buf[i]) * 1099511628211ull;
fclose(f);
return total > 12 ? (h | 1) : 0;
}
/// Whether the shipped list is one the notes were not made from yet
bool state_shipped_new() {
if (getenv("RX_VK_NO_STATE_LIST")) return false;
const uint64_t id = state_shipped_id();
if (!id) return false;
uint64_t had = 0;
if (FILE* f = fopen(state_list_path("triforce_states_applied.bin").c_str(), "rb")) {
if (fread(&had, 8, 1, f) != 1) had = 0;
fclose(f);
}
return had != id;
}
/// (prepare_run, the seed open) every pipeline of the state lists (the player's own and the shipped one): its program translated
/// when the program cache lacks it, then noted in triforce_pipelines.bin for precompile_run to build. RX_VK_NO_STATE_LIST=1: off
void state_list_prepare(void (*progress)(size_t done, size_t total)) {
if (getenv("RX_VK_NO_STATE_LIST")) return;
const auto t0 = std::chrono::steady_clock::now();
const uint64_t shipped_id = state_shipped_id();
struct Applied { // (written when the lists were gone through, however this returns)
uint64_t id;
~Applied() {
if (!id) return;
if (FILE* f = fopen(state_list_path("triforce_states_applied.bin").c_str(), "wb")) fwrite(&id, 8, 1, f), fclose(f);
}
} applied{shipped_id};
std::vector<StateRec> recs = state_list_read(state_list_path("triforce_pipeline_states_shipped.bin"));
const size_t shipped = recs.size();
for (const StateRec& r : state_list_read(state_list_path("triforce_pipeline_states.bin"))) recs.push_back(r);
@@ -1220,6 +1255,9 @@ void state_list_prepare(void (*progress)(size_t done, size_t total)) {
rec_program[i] = it->second;
}
std::atomic<size_t> next{0}, made{0}, missing{0};
// (perf push 11-11: why a program of the list has no key. 0 a header-only stage, 1 mesh on a GPU without, 2 a stage the
// seed lacks or has as another kind. Measured on OoT: all 2178 of 6150 are the last, vertex and fragment both absent)
std::atomic<size_t> why_missing[3]{};
const auto worker = [&] {
ub_thread_off(); // (a background thread: see ub_thread_off; the pipeline state list's builds raced the replay's updates)
SetThreadPriority(GetCurrentThread(), THREAD_PRIORITY_BELOW_NORMAL);
@@ -1230,7 +1268,15 @@ void state_list_prepare(void (*progress)(size_t done, size_t total)) {
std::unique_ptr<Program> p;
// (a mesh program on a GPU without mesh shaders is not made; one with a header-only stage is made in play)
if (!r.code[3] && (!r.mesh || dev.mesh_shader)) p = seed_program_of(r.code, r.mesh != 0, r.vattr, r.color_types, &want[i].key);
if (!want[i].key) missing++;
if (!want[i].key) {
missing++;
why_missing[r.code[3] ? 0 : r.mesh && !dev.mesh_shader ? 1 : 2]++;
static std::atomic<int> told{0};
if (getenv("RX_VK_SEED_CHECK") && told++ < 40)
logf("vk: state list: no key for %016llx %016llx %016llx %016llx mesh %u",
(unsigned long long)r.code[0], (unsigned long long)r.code[1], (unsigned long long)r.code[2],
(unsigned long long)r.code[3], r.mesh);
}
if (p) {
if (!p->failed) made++;
for (Stage& st : p->stages)
@@ -1311,6 +1357,10 @@ void state_list_prepare(void (*progress)(size_t done, size_t total)) {
"by their textures (%zu made), %zu noted, %zu whose program variant was not made, in %.1f s", recs.size(), shipped,
want.size(), made.load(), missing.load(), variant_of.size(), variants_made.load(), noted_n, no_variant,
std::chrono::duration<double>(std::chrono::steady_clock::now() - t0).count());
if (missing.load())
logf("vk: state list: the %zu programs without a key: %zu with a header-only stage, %zu mesh (none on this GPU), %zu with a "
"stage the game's shader list lacks (made in play)", missing.load(), why_missing[0].load(), why_missing[1].load(),
why_missing[2].load());
}
// ---- (hookshot) Stage libraries, as TotK Recomp's runtime/vk_pipeline.c: the part of a pipeline that costs is its shaders, and with
@@ -2060,8 +2110,10 @@ struct HandleHit {
uint64_t destroy_gen;
bool sampling; // (its use reads the texture's contents)
uint64_t tex_epoch; // (the texture epoch it was last touched in)
bool known = true; // (Triforce: the slot is registered: with no view, a texture whose image is not there yet)
};
TLS_OBJECT(handle_hits, std::unordered_map<uint64_t, HandleHit>)
thread_local bool view_slot_known = true; // (Triforce: the last texture_view_of's slot was a registered one)
// (hookshot) RX_VK_TRACE_TEX=<w>x<h>,...: each draw sampling a texture of that size is logged with its state, every 30th frame, 600 lines: a
// model that does not show is found by its textures, and whether the game draws it, and how, read from the line
thread_local const Texture* trace_tex_hit = nullptr;
@@ -2125,6 +2177,7 @@ VkImageView texture_view_of(uint32_t handle, VkImageViewType want, ViewUse use,
}
if (tex_out) *tex_out = it->second.tex;
if (level_out) *level_out = it->second.level;
view_slot_known = it->second.known;
return it->second.view;
}
ParLock lk;
@@ -2141,6 +2194,31 @@ VkImageView texture_view_of(uint32_t handle, VkImageViewType want, ViewUse use,
if (tex) level = image_level(tex, has_view ? view_level(&v) : 0);
}
if (!view) { // (what a shader reads as the dummy instead: once per texture slot)
{ // (Triforce, diagnostics) RX_VK_LOG_DUMMY=1: every 10 s, the slots that were resolved to the dummy in the last second
static const bool log_dummy = getenv("RX_VK_LOG_DUMMY") != nullptr;
if (log_dummy) {
struct Seen { int64_t n = 0; std::chrono::steady_clock::time_point last; char desc[200]; };
static std::unordered_map<uint32_t, Seen> seen;
static auto next = std::chrono::steady_clock::now() + std::chrono::seconds(10);
const auto now = std::chrono::steady_clock::now();
Seen& d = seen[handle & 0xFFFFF];
d.n++, d.last = now;
if (!known) snprintf(d.desc, sizeof d.desc, "not registered");
else snprintf(d.desc, sizeof d.desc, "%s; target %d format %#x %ux%ux%u levels %u gpu %#llx view %s (format %#x target %d) want %d use %d",
tex ? "no view of it" : "no image for it", t.d.target, t.d.format, t.d.width, t.d.height, t.d.depth, t.d.levels,
(unsigned long long)t.gpu, has_view ? "yes" : "no", has_view && v.set_format ? v.format : 0,
has_view && v.set_target ? v.target : -1, (int)want, (int)use);
if (now >= next) {
next = now + std::chrono::seconds(10);
int k = 0;
for (auto& [slot, e] : seen)
if (now - e.last < std::chrono::seconds(1) && k++ < 48)
logf("vk: dummy: slot %#x resolved %lld times in 10 s: %s", slot, (long long)e.n, e.desc);
logf("vk: dummy: %d slots resolved to the dummy in the last second", k);
for (auto& [slot, e] : seen) e.n = 0;
}
}
}
static std::unordered_map<uint32_t, int> told;
if (told.size() < 64 && !told[handle & 0xFFFFF]++) {
if (!known)
@@ -2153,7 +2231,8 @@ VkImageView texture_view_of(uint32_t handle, VkImageViewType want, ViewUse use,
has_view && v.set_format ? v.format : 0, has_view && v.set_target ? v.target : -1, (int)want, (int)use);
}
}
handle_hits()[key] = {view, tex, level, version, dgen, use == ViewUse::Sample || use == ViewUse::Storage, texture_epoch()};
handle_hits()[key] = {view, tex, level, version, dgen, use == ViewUse::Sample || use == ViewUse::Storage, texture_epoch(), known};
view_slot_known = known;
if (tex_out) *tex_out = tex;
if (level_out) *level_out = level;
return view;
@@ -2328,6 +2407,7 @@ bool dump_sampled_wanted(uint64_t key) {
} // (RX_VK_DUMP: the storage images the descriptors let the program write)
TLS_OBJECT(dump_inputs, std::string) // (and what it samples)
int64_t st_dummy_textures; // (textures sampled as the dummy: no image for the handle)
int64_t st_dummy_known; // (Triforce: of those, registered textures: their image not made or not uploaded yet)
void dump_input(char kind, uint32_t index, uint32_t handle, const Texture* tex) {
char b[160];
@@ -2877,7 +2957,7 @@ void build_descriptors(const NvnState& s, const Program& p) {
tex ? tex->width : 0, tex ? tex->height : 0, tex ? tex->layers : 0, tex ? (void*)tex->image : nullptr,
(void*)view, view ? "" : " (dummy)");
if (view && tex && tex->gpu_written) draw_sampled().push_back(tex);
if (!view) view = dummies.view[(int)vt], st_dummy_textures += handle != 0;
if (!view) view = dummies.view[(int)vt], st_dummy_textures += handle != 0, st_dummy_known += handle != 0 && view_slot_known;
dw().imgs.push_back({sampler_for((int)(handle >> 20), d.is_depth), view, VK_IMAGE_LAYOUT_GENERAL});
}
w.pImageInfo = &dw().imgs[first];
@@ -2942,6 +3022,10 @@ TLS_OBJECT(pushed, std::array<Pushed, 2>) // [graphics, compute]
/// to it)
VkPipelineCache shared_pipeline_cache() { return pcache; }
void draw_skip_report(double frames) { skip_report_impl(frames), fine_report(frames); }
/// (Triforce) how many texture bindings have sampled the dummy so far (a texture with no image yet: left for a later frame, not
/// made, evicted): RX_VK_FRAME_TRACE's last column, the measure of materials showing flat while their textures come in
/// Registered textures only: OoT binds one handle of no registered slot (0x3126f) in ~18 places a frame, always the dummy
int64_t draw_dummy_textures() { return st_dummy_known; }
void descriptors_disturbed() {
pushed()[0].cmd = pushed()[1].cmd = VK_NULL_HANDLE;
}
@@ -7102,6 +7186,15 @@ bool prepare_run(void (*progress)(size_t done, size_t total), bool background) {
logf("vk: the shader list was hashed (%.0f MB in %.0f ms): written since the last check, or too close to it to go by its time",
prepared_for.seed_size / 1048576.0, hash_ms);
if (why == PrepareWhy::None && !getenv("RX_VK_PREPARE_AGAIN")) {
// (perf push 11-11: a shipped state list the notes were not made from, in a folder prepared already: its pipelines are
// noted now, for the precompile that follows, as a prepare would have. See state_shipped_id)
if (!background && state_shipped_new()) {
logf("vk: the shipped pipeline list is not the one this folder's notes were made from: its pipelines are noted now");
if (seed_open()) {
state_list_prepare(nullptr);
seed_close();
}
}
if (getenv("RX_VK_SEED_CHECK")) seed_open();
if (prepared_for.warmed == warmed || background) return false;
// (translated, and the driver's part is another driver's, or of before the libraries were wanted: that part alone, no
@@ -7279,6 +7372,47 @@ void prepare_background() {
}).detach();
}
// ---- (perf push 11-11) The precompile on a PC with little RAM. Every pipeline built here is kept for the session, with its
// program: measured at a fresh Market (RTX 4070 SUPER), the process holds 12.9 GB in RAM without the precompile, 13.4 GB with
// 4704 graphics and 275 compute pipelines, 15.7 GB with 8550 and 1650; after a tour of ten areas 18.6 against 20.6 GB. What
// the precompile is for is the compile, which the caches keep (ours, mkw_vk_pipelines.bin, and the driver's): with them warm,
// the same runs without it built 1462 and 4046 pipelines in play in 357 and 1639 ms in all (0.2-0.4 ms each), with as many
// frames over 40, 100 and 250 ms as with it. So where RAM is short (under 20 GB in the PC: the 16 GB ones, whose logs show the
// system taking pages back all through play) a pipeline is built before the game once, on the first start that has its note,
// and that start records it (triforce_pipelines.bin.built, for this translator and driver); later starts leave it to its first
// draw, and neither it nor its program is in RAM until then.
// The price, same tour: of ~4050 pipelines then built in play, ten or twelve submissions spend 10-65 ms on them (some in frames
// that were slow anyway, an area's arrival; three or four make a 40 ms frame of their own), and one in two tours had a
// pipeline the caches missed (508 ms). So it is not the default: RX_VK_PRECOMPILE_KEEP=auto: this rule where the PC has under
// 20 GB of RAM; =0: on any PC; unset or 1: every start builds and keeps all, as before ----
constexpr uint64_t kBuiltComputes = 0xC04D50555445D001ull; // (in the record: the compute programs' pipelines were built)
bool precompile_keeps_all() {
const char* e = getenv("RX_VK_PRECOMPILE_KEEP");
if (!e) return true;
if (strcmp(e, "auto")) return atoi(e) != 0;
#ifdef _WIN32
MEMORYSTATUSEX ms{};
ms.dwLength = sizeof ms;
if (GlobalMemoryStatusEx(&ms)) return ms.ullTotalPhys >= (20ull << 30);
#endif
return true;
}
std::unordered_set<uint64_t> built_read(uint64_t stamp) {
std::unordered_set<uint64_t> keys;
if (FILE* f = fopen((pipe_notes_path() + ".built").c_str(), "rb")) {
uint64_t v = 0;
if (fread(&v, 8, 1, f) == 1 && v == stamp)
while (fread(&v, 8, 1, f) == 1) keys.insert(v);
fclose(f);
}
return keys;
}
void built_write(uint64_t stamp, const std::unordered_set<uint64_t>& keys) {
std::vector<uint64_t> all{stamp};
all.insert(all.end(), keys.begin(), keys.end());
if (FILE* f = fopen((pipe_notes_path() + ".built").c_str(), "wb")) fwrite(all.data(), 8, all.size(), f), fclose(f);
}
void precompile_run(void (*progress)(size_t done, size_t total)) {
const auto t0 = std::chrono::steady_clock::now();
// (a start that dies in here would die in here every time, the notes being what it builds from: a player saw the console reach
@@ -7312,12 +7446,29 @@ void precompile_run(void (*progress)(size_t done, size_t total)) {
std::lock_guard lk(state_mu);
if (!state_file) state_open_locked();
}
// (little RAM, see precompile_keeps_all: the pipelines an earlier start built for these caches are left to their first draw)
const bool light = !precompile_keeps_all();
const uint64_t built_stamp = warmed_stamp(gpl_wanted());
std::unordered_set<uint64_t> built;
const size_t noted_n = notes.size();
bool computes_built = false;
if (light) {
built = built_read(built_stamp);
computes_built = built.count(kBuiltComputes) != 0;
std::erase_if(notes, [&](const std::unique_ptr<PipeNote>& n) { return built.count(n->key) != 0; });
logf("vk: precompile: RAM spared (RX_VK_PRECOMPILE_KEEP=1: as before): %zu of %zu noted pipelines were built by "
"an earlier start and are left to their first draw%s", noted_n - notes.size(), noted_n,
computes_built ? ", the compute programs' too" : "");
}
std::vector<uint64_t> building; // (their keys, for the record once they are built)
if (light)
for (const auto& n : notes) building.push_back(n->key);
std::unordered_set<uint64_t> wanted;
for (const auto& n : notes) wanted.insert(n->prog);
// (the compute programs: those play dispatched (triforce_computes.bin), or every kept one when there is no list yet)
const char* which = getenv("RX_VK_PRECOMPILE_COMPUTES");
const std::unordered_set<uint64_t> used = computes_read();
const bool all_computes = which ? !strcmp(which, "all") : used.empty();
const std::unordered_set<uint64_t> used = computes_built ? std::unordered_set<uint64_t>{} : computes_read();
const bool all_computes = !computes_built && (which ? !strcmp(which, "all") : used.empty());
if (!all_computes && !(which && !strcmp(which, "none"))) wanted.insert(used.begin(), used.end());
std::vector<Program*> computes, fallbacks;
const size_t programs_n = program_cache_load_some(wanted, [&](Program& p) {
@@ -7431,6 +7582,13 @@ void precompile_run(void (*progress)(size_t done, size_t total)) {
} else {
for (std::thread& t : threads) t.join();
delete pre;
// (little RAM: what this start built is on disk before the record says so, and the next start leaves it out)
if (light && !prepare_stopped() && (!building.empty() || !computes_built)) {
built.insert(building.begin(), building.end());
built.insert(kBuiltComputes);
pipeline_cache_save(true);
built_write(built_stamp, built);
}
}
logf("vk: precompile: %zu programs, %zu graphics and %zu compute pipelines, %zu mesh fallback programs in %.0f ms (%.0f ms "
"of it the caches opened, the notes and programs read)", programs_n, graphics_n, computes_n, fallbacks_n,

View File

@@ -65,6 +65,7 @@ void dry_marks_begin(); // (at the start of each dry pass: the last targets it
void draw_stats(int64_t* draws, int64_t* skipped, int64_t* dispatches, int64_t* pipelines);
void draw_stats2(int64_t* passes, int64_t* mesh, double* pipeline_ms);
void draw_stats3(int64_t* pending_skips, int64_t* waits, double* wait_ms);
int64_t draw_dummy_textures(); // (Triforce: texture bindings that sampled the dummy so far: RX_VK_FRAME_TRACE)
/// (hookshot) pipelines linked from stage libraries in the draw, and the libraries asked of the driver for them: had, and cold;
/// of the cold, those a worker has built since, and the workers' time in them. False: no stage libraries (RX_VK_GPL)
bool draw_stats4(int64_t* links, double* link_ms, int64_t* lib_hits, int64_t* lib_cold, int64_t* lib_built, double* lib_build_ms);

View File

@@ -210,6 +210,9 @@ void memory_drop_shadow(uint64_t guest, uint64_t size); // (a write the captur
/// when any write-back ever did
uint64_t memory_writeback_value(uint64_t guest, uint64_t size, bool* ever);
void memory_note_readback(uint64_t guest, uint64_t size); // (PR 430) a CPU read of GPU results: see memory_rb_refresh
/// (Triforce) around the write of an NVN event's word by a command buffer's signal: its write-back is not one the captures of
/// its page wait for (vk_memory.cpp). This thread's bindings while on
void memory_bind_event_write(bool on);
bool memory_rb_refresh(VkCommandBuffer up);
void memory_rb_recording_begun(); // (hookshot) at a recording's start, on its thread: the submission it joins (memory_rb_refresh)
void memory_rb_segments_begin(uint64_t base, uint64_t n); // (and a parallel replay's segments: their order, base on)
@@ -219,11 +222,12 @@ void memory_note_gpu_write(uint64_t gpu, uint64_t size);
uint64_t vk_timeline_completed();
void vk_timeline_wait(uint64_t v);
uint32_t memory_generation();
uintptr_t memory_stamp_who(uint64_t guest, uint32_t* gen, uintptr_t* anchor); // (RX_VK_STAMP_SRC: who wrote that page last)
/// 1: a page of [guest, guest + size) was written after generation gen; 0: none was
bool memory_written_since(uint64_t guest, uint64_t size, uint32_t gen);
void memory_stamp(uint64_t guest, uint64_t size); // a write the watch does not see (GPU writes, host writes)
/// The page runs of [guest, guest + size) written after gen, as byte ranges from guest (appended to out)
void memory_written_runs(uint64_t guest, uint64_t size, uint32_t gen, std::vector<std::pair<uint64_t, uint64_t>>& out);
void memory_written_runs(uint64_t guest, uint64_t size, uint32_t gen, std::vector<std::pair<uint64_t, uint64_t>>& out, bool tex = false);
uint64_t memory_pages_collected(); // (statistics)
/// An NVN format as Vulkan stores it (vk_formats.cpp): ASTC is decoded to RGBA8, block_w x block_h
@@ -271,6 +275,19 @@ bool create_buffer(uint64_t size, VkBufferUsageFlags usage, bool host_visible, V
void destroy_buffer(VkBuffer buffer, VmaAllocation allocation);
/// A buffer in VRAM only (within the driver's budget), or false
bool create_device_buffer(uint64_t size, VkBufferUsageFlags usage, VkBuffer& buffer, VmaAllocation& allocation, bool dedicated = false);
/// (Triforce) device memory for a piece of a sparse buffer, and its release (vk_vma.cpp)
bool vma_piece_alloc(uint64_t size, uint64_t alignment, uint32_t type_bits, VmaAllocation& allocation, VkDeviceMemory& memory,
uint64_t& offset);
void vma_piece_free(VmaAllocation allocation);
/// (Triforce) the mirrors' pieces made resident in this recording, bound before it is submitted (vk_memory.cpp)
void memory_mirror_binds_flush();
bool memory_mirror_bind_wait(VkSemaphoreSubmitInfo* wait); // (and the wait a queue submission makes for those binds)
/// (Triforce, vk_memory.cpp) The memory handling played on NVIDIA cards only (mirrors in pieces, compacted textures): on there,
/// and on trial on a card of another make until triforce_memtrial.txt says it passed or failed. tick: at the 600th presents;
/// device_lost: from device_lost()
bool memory_trial_on();
void memory_trial_tick();
void memory_trial_device_lost();
/// Device-local memory this process uses and what the driver allows it now (VK_EXT_memory_budget), summed over the
/// device-local heaps; false when unknown
bool vram_usage(uint64_t* used, uint64_t* budget, uint64_t* block_free = nullptr);

View File

@@ -885,6 +885,7 @@ void prepare_with_screen() {
std::thread worker([] {
const bool update = prepare_is_update(); // (an update, RX_VK_PREPARE_BACKGROUND=1: what the notes name now, the rest in play)
const bool prepared = !update && prepare_run(prep_progress, false);
if (update) state_list_run(); // (as the start without a screen does, below: the state lists' pipelines before the precompile)
if (prepared && !prepare_stopped()) { // (a first start only: the pipelines the notes predict, into the driver's cache)
prep_done.store(0), prep_total.store(0);
prep_stage.store(1);

File diff suppressed because it is too large Load Diff

View File

@@ -180,6 +180,8 @@ void queue_flush_locked() { // (dev.present_mu held)
si.pCommandBufferInfos = pend.cbs.data();
si.signalSemaphoreInfoCount = n;
si.pSignalSemaphoreInfos = sig;
VkSemaphoreSubmitInfo bind_wait; // (Triforce: after the memory of the mirrors' new pieces is bound, vk_memory.cpp)
if (memory_mirror_bind_wait(&bind_wait)) si.waitSemaphoreInfoCount = 1, si.pWaitSemaphoreInfos = &bind_wait;
PROF(PROF_QSUBMIT);
const VkResult res = vkQueueSubmit2(dev.queue, 1, &si, VK_NULL_HANDLE);
static int failures;
@@ -355,6 +357,8 @@ void submit_recording(Recording& r, bool signal_value = true, uint64_t signal_at
return;
}
PROF(PROF_QSUBMIT);
VkSemaphoreSubmitInfo bind_wait; // (Triforce: after the memory of the mirrors' new pieces is bound, vk_memory.cpp)
if (memory_mirror_bind_wait(&bind_wait)) si.waitSemaphoreInfoCount = 1, si.pWaitSemaphoreInfos = &bind_wait;
const VkResult res = vkQueueSubmit2(dev.queue, 1, &si, r.fence);
r.submitted = res == VK_SUCCESS;
static int failures;
@@ -732,7 +736,9 @@ void replay(VkCommandBuffer& cmd, NvnState& s, const uint8_t* p, const uint8_t*
end_pass(cmd);
GprofSpan gp(cmd, "event signal (barrier, write, barrier)");
full_barrier(cmd);
memory_bind_event_write(true); // (Triforce: not a write-back the captures of its page wait for)
gpu_write(cmd, e.gpu, &v, 4);
memory_bind_event_write(false);
break;
}
case NOP_WAIT_EVENT: { // (a wait for an event another queue's work signals: submit_job orders them)
@@ -1282,10 +1288,76 @@ void maybe_split(VkCommandBuffer& cmd) {
// (hookshot) One cause of those is gone: the dry pass did not mark the targets of a clear (NOP_CLEAR_COLOR in replay()). With
// that, Castle Town's market and eight other places showed no black frame on 3 to 8 threads; the Kakariko walk was not run
// again, so it stays off. On a 9950X3D the replay's wall time in the market goes from 10.9 ms a frame to 7.75 on 3 threads.
int par_threads() {
static const int n = std::clamp(getenv("RX_VK_PARALLEL") ? atoi(getenv("RX_VK_PARALLEL")) : 0, 0, 8);
// (Triforce) RX_VK_PARALLEL_AUTO=<n> (2 to 8): the parallel replay on <n> threads where it is the faster of the two, measured
// as the game runs. On three threads it is faster where the replay is the frame's limit and slower where the GPU is: at the
// Castle Town save 67.1 fps against 58.8 (RTX 4070 SUPER, 2560 x 1440, windowed), on a tour of Hyrule Field, Kakariko and
// Lake Hylia 59.5 and 61.9 against 64.8 and 66.7 (its recordings done sooner, the game's thread only waits longer for the
// GPU, and the dry pass and the hand-over are not free); the flicker counts of the frame log were the same either way.
// par_tune_present tries the other way for a moment every so often and keeps the one whose frames are shorter.
// RX_VK_PARALLEL=<n> still means always, as before.
// EXPERIMENTAL, off unless asked for: its first runs (2026-11-11) gained nothing, 54.7 fps against 58.9 standing in Castle Town
// and 60.4 against 61.7 on the tour. The frame time of a scene with a crowd moves by 2 ms within seconds (one trial read 18.59
// and then 16.44 ms for the same way), more than the difference being measured, so the trials decide on noise
int par_threads_max() {
static const int n = [] {
if (const char* e = getenv("RX_VK_PARALLEL")) return std::clamp(atoi(e), 0, 8);
if (const char* e = getenv("RX_VK_PARALLEL_AUTO")) return std::clamp(atoi(e), 0, 8);
return 0;
}();
return n;
}
bool par_auto() {
static const bool v = !getenv("RX_VK_PARALLEL") && par_threads_max() >= 2;
return v;
}
bool par_auto_on = false; // (the replay thread's: the parallel replay is the one in use now)
int par_threads() {
const int n = par_threads_max();
return par_auto() && !par_auto_on ? 0 : n;
}
/// At each present (the replay thread): the tuner. Every kEvery presents it times kTake frames as they are, kTake the other way
/// (after kSettle to settle) and kTake as they were again, and changes over when the other way's mean frame was at least 4%
/// shorter than both: a scene that changed under the trial does not pass for a gain. Frames over 60 ms (a load) are left out
void par_tune_present() {
if (!par_auto()) return;
constexpr int kEvery = 900, kTake = 80, kSettle = 24;
static auto last = std::chrono::steady_clock::now();
const auto now = std::chrono::steady_clock::now();
const double ms = std::chrono::duration<double, std::milli>(now - last).count();
last = now;
static int phase = 0, n = 0, cnt[3];
static double sum[3];
static int64_t frames = 0, next_trial = 600, st_trials = 0, st_changes = 0;
frames++;
if (phase == 0) {
if (frames < next_trial) return;
phase = 1, n = 0;
sum[0] = sum[1] = sum[2] = 0, cnt[0] = cnt[1] = cnt[2] = 0;
return;
}
const int k = phase - 1;
n++;
if (n > (phase == 1 ? 0 : kSettle) && ms <= 60.0) sum[k] += ms, cnt[k]++;
if (n < (phase == 1 ? kTake : kSettle + kTake)) return;
n = 0;
if (phase == 1) {
par_auto_on = !par_auto_on, phase = 2;
} else if (phase == 2) {
par_auto_on = !par_auto_on, phase = 3;
} else {
phase = 0, st_trials++;
const bool ok = cnt[0] >= kTake * 3 / 4 && cnt[1] >= kTake * 3 / 4 && cnt[2] >= kTake * 3 / 4;
const double a = cnt[0] ? sum[0] / cnt[0] : 0, b = cnt[1] ? sum[1] / cnt[1] : 0, c = cnt[2] ? sum[2] / cnt[2] : 0;
const bool better = ok && b < std::min(a, c) * 0.96;
if (better) par_auto_on = !par_auto_on, st_changes++;
next_trial = frames + (better ? kEvery : kEvery * 2);
if (better || st_trials <= 8 || st_trials % 20 == 0)
logf("vk: parallel replay, trial %lld: %s %.2f ms a frame, the other way %.2f, %s again %.2f: %s (%lld changes so far)",
(long long)st_trials, par_auto_on != better ? "parallel" : "serial", a, b, par_auto_on != better ? "parallel" : "serial", c,
better ? (par_auto_on ? "parallel from now on" : "serial from now on") : (ok ? "as it was" : "not measured (a load)"),
(long long)st_changes);
}
}
struct ParPool {
std::mutex mu;
std::condition_variable cv, done;
@@ -1531,6 +1603,8 @@ const uint8_t* replay_handle(VkCommandBuffer& cmd, const uint8_t* h, bool may_st
si.pCommandBufferInfos = cb;
std::lock_guard lk(dev.present_mu);
queue_flush_locked();
VkSemaphoreSubmitInfo bind_wait; // (Triforce: after the memory of the mirrors' new pieces is bound, vk_memory.cpp)
if (memory_mirror_bind_wait(&bind_wait)) si.waitSemaphoreInfoCount = 1, si.pWaitSemaphoreInfos = &bind_wait;
const VkResult res = vkQueueSubmit2(dev.queue, 1, &si, j->r->fence);
j->r->submitted = res == VK_SUCCESS;
if (res != VK_SUCCESS) logf("vk: segment submit %llu failed (%d)", (unsigned long long)j->r->value, (int)res);
@@ -2559,6 +2633,7 @@ uint32_t crc32(const uint8_t* p, size_t n, uint32_t c = 0xFFFFFFFFu) {
/// closed (issue 1275: captures kept 25 -> 6527, staging 256 -> 1408 MB). Now the report once (a thread that sees the loss
/// too waits for it, instead of ending the process before it is written), then the process ends as replay_fatal does
[[noreturn]] void device_lost() {
memory_trial_device_lost(); // (Triforce: with the memory handling on trial on this card: off from the next start)
static std::once_flag once;
std::call_once(once, checkpoints_report);
replay_fatal("the GPU device was lost (VK_ERROR_DEVICE_LOST): the driver stopped or reset the GPU, and nothing sent "
@@ -2837,7 +2912,8 @@ void status_tick() {
std::string dir(exe, n);
FILE* f = fopen((dir.substr(0, dir.find_last_of("\\/") + 1) + "triforce_frames.csv").c_str(), "w");
if (f) fprintf(f, "frame,ms,pipe_waits,pipe_wait_ms,left_out,translate_ms,uploads,upload_mb,textures,memsync_ms,pipelines_ms,"
"uploads_ms,programs_ms,slot_wait_ms,gpu_behind_ms,draw_ms,desc_ms,cmds_ms,tex_lookup_ms,textures_gone\n");
"uploads_ms,programs_ms,slot_wait_ms,gpu_behind_ms,draw_ms,desc_ms,cmds_ms,tex_lookup_ms,textures_gone,"
"dummies,draws\n");
return f;
}();
if (trace && last_ns) {
@@ -2851,12 +2927,17 @@ void status_tick() {
const auto dp = [](int i) { return (prof_ns[i] - l_prof[i]) / 1e6; };
static uint64_t l_gone;
const uint64_t gone = texture_destroy_generation();
fprintf(trace, "%llu,%.2f,%lld,%.2f,%lld,%.1f,%lld,%.2f,%zu,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%llu\n",
static int64_t l_dummy, l_draws; // (Triforce: the bindings that sampled the dummy, and the draws, this frame)
const int64_t dummy = draw_dummy_textures();
int64_t draws_now = 0, sk_now = 0, di_now = 0, pl_now = 0;
draw_stats(&draws_now, &sk_now, &di_now, &pl_now);
fprintf(trace, "%llu,%.2f,%lld,%.2f,%lld,%.1f,%lld,%.2f,%zu,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%.2f,%llu,%lld,%lld\n",
(unsigned long long)frames, (now_ns - last_ns) / 1e6, (long long)(pw - l_pw), pwms - l_pwms, (long long)(ps - l_ps),
tr - l_tr, (long long)(up - l_up), (upb - l_upb) / 1048576.0, texture_count(), dp(PROF_MEMSYNC), dp(PROF_PIPE),
dp(PROF_UPLOAD), dp(PROF_PROGRAM), dp(PROF_SLOTWAIT), (st_wait_ns - l_wait) / 1e6, dp(PROF_DRAW), dp(PROF_DESC),
dp(PROF_CMDS), dp(PROF_TEXGET), (unsigned long long)(gone - l_gone));
l_gone = gone;
dp(PROF_CMDS), dp(PROF_TEXGET), (unsigned long long)(gone - l_gone), (long long)(dummy - l_dummy),
(long long)(draws_now - l_draws));
l_gone = gone, l_dummy = dummy, l_draws = draws_now;
if (frames % 60 == 0) fflush(trace);
l_ps = ps, l_pw = pw, l_pwms = pwms, l_up = up, l_upb = upb, l_tr = tr, l_wait = st_wait_ns;
std::memcpy(l_prof, prof_ns, sizeof l_prof);
@@ -2964,6 +3045,7 @@ void do_present(const nvn_tex* t, const int crop[4]) {
full_barrier(r.cmd);
presents++;
presents_now = presents;
par_tune_present(); // (Triforce: RX_VK_PARALLEL_AUTO)
wide_sides_tick();
if (presents == 1) hang_watch_start();
{ // (EvilBoyce, diagnostics) RX_VK_WATCH_CPU=<present>:<gpu address>[,...] (hex addresses): from that present on, the
@@ -4178,7 +4260,7 @@ bool replay_start() {
VkSemaphoreCreateInfo si{VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO};
vkCreateSemaphore(dev.device, &si, nullptr, &f.acquired);
}
if (par_threads() >= 2 && !init_pools_make(pi)) return false; // (last: nothing after it fails and leaves them made)
if (par_threads_max() >= 2 && !init_pools_make(pi)) return false; // (last: nothing after it fails and leaves them made)
logf("vk: replay ready (%s; push descriptors %s (%u), extended dynamic state 3 %s, custom border colours %s)",
dev.offscreen ? "offscreen" : "window", dev.push_descriptor ? "yes" : "no", dev.max_push_descriptors,
dev.eds3_blend ? "yes" : "no", dev.custom_border_color ? "yes" : "no");

View File

@@ -241,7 +241,21 @@ std::unordered_map<uint64_t, std::vector<Pooled>> target_pool; // by pool_key_
uint64_t pool_bytes, pool_n, st_pool_reused, st_pool_made, st_pool_destroyed;
bool pool_pressed; // (VRAM near what the driver allows: texture_frame's budget)
bool compact_on() { // (hookshot, RX_VK_COMPACT=1: the textures in a pool of their own, compacted: compact_step)
static const bool v = getenv("RX_VK_COMPACT") && atoi(getenv("RX_VK_COMPACT")) != 0;
// (Triforce) on by default on NVIDIA cards (RX_VK_COMPACT=0: off; =1: on, any card). Measured there: an RTX 4070 SUPER
// played with it since 2026-11-10 (a 584,000-present session, 132 compactions, no fault; 85 -> 80.6 fps after a six-area
// tour against 81 -> 64 without), and with the allowance held to 7300 MB (RX_VK_TEST_ALLOWED_MB, an 8 GB card's) a tour of
// Hyrule Field, Kakariko, Lake Hylia and Kokiri peaked at 7165 MB of VRAM against 8565 without (over the allowance in 2 of
// 12 reports against 11) at 69 fps against ~64. On an 8 GB card what is over the allowance is paged to system memory, and
// the textures and render targets evicted to get back under it are made and uploaded again, 58 to 190 ms a frame (an
// RX 5700 XT's and an RTX 3060's logs, 2026-11-11). Not measured on AMD or Intel: as it was there until someone does
// (Triforce, later that day) and on trial on the other makes' cards: memory_trial_on (vk_memory.cpp), which turns it off by
// itself on a PC where a session with it loses the GPU or two end in a crash
static const bool v = [] {
if (const char* e = getenv("RX_VK_COMPACT")) return atoi(e) != 0;
const bool on = memory_trial_on();
if (on) logf("vk: the textures' VRAM is compacted (RX_VK_COMPACT=0: off)");
return on;
}();
return v;
}
uint64_t pool_cap() {
@@ -1297,6 +1311,7 @@ struct UpSlow { int64_t ns; uint32_t w, h, layers, nvn; int route; uint64_t byte
std::vector<UpSlow> up_slowest;
std::mutex up_mu; // (uploads may come from a parallel replay's threads)
std::atomic<int64_t> frame_upload_ns{0}; // (this frame's time in upload(), every thread's)
std::atomic<int> frame_astc_admitted{0}; // (Triforce: this frame's new ASTC textures asked about so far: upload_defer)
std::atomic<int64_t> st_deferred{0}, st_deferred_frame{0};
void upload_frame_report() {
std::lock_guard lk(up_mu);
@@ -1589,11 +1604,13 @@ bool upload_inner(Texture& tex, const Ranges* dirty) {
/// What the CPU wrote of a texture's storage since gen, as byte ranges from its start. A texture in a virtual
/// (sparse) pool is looked at page by page through its mapping, and is all dirty when its mapping changed.
bool texture_dirty(Texture& tex, uint32_t gen, Ranges& out) {
/// same_ok (Triforce): a capture's stamp of a page with the bytes of its last copy is no write (vk_memory.cpp page_gen_tex): only
/// for an uploaded texture's "written since my upload?"
bool texture_dirty(Texture& tex, uint32_t gen, Ranges& out, bool same_ok = false) {
out.clear();
constexpr uint64_t kVirtual = 1ull << 40, kVPage = 1ull << 16;
if (tex.gpu < kVirtual) {
memory_written_runs(tex.guest, tex.guest_size, gen, out);
memory_written_runs(tex.guest, tex.guest_size, gen, out, same_ok);
return !out.empty();
}
// (the page walk only when the GPU mappings changed: the terrain's sparse arrays span over 1 GB, 64 KB pages,
@@ -1661,7 +1678,7 @@ bool texture_dirty(Texture& tex, uint32_t gen, Ranges& out) {
const size_t remaps = out.size();
for (const Texture::Run& r : tex.runs) {
in_run.clear();
memory_written_runs(r.guest, r.n, gen, in_run);
memory_written_runs(r.guest, r.n, gen, in_run, same_ok);
for (const auto& [lo, hi] : in_run) {
if (out.size() > remaps && out.back().second == r.off + lo) out.back().second = r.off + hi;
else out.push_back({r.off + lo, r.off + hi});
@@ -1693,6 +1710,14 @@ void destroy_later(Texture& tex) {
}
d.memory = tex.sparse_chunks;
dead.push_back(std::move(d));
if (tex.twin3d) { // (Triforce: its one-slice 3D copy goes with it)
Dead t{tex.last_value, tex.twin3d, tex.twin3d_alloc, {}, {}};
for (auto& kv : tex.twin3d_views)
if (kv.second) t.views.push_back(kv.second);
dead.push_back(std::move(t));
tex.twin3d = VK_NULL_HANDLE, tex.twin3d_alloc = nullptr;
tex.twin3d_views.clear();
}
total_bytes -= std::min(total_bytes, tex.bytes + tex.pad);
}
@@ -2175,6 +2200,7 @@ VkCommandBuffer upload_command_buffer() {
}
bool texture_uploads_recorded() {
if (memory_rb_refresh(up_cmd)) up_used = true; // (PR 430: readback pages from guest memory, last of the uploads)
memory_mirror_binds_flush(); // (Triforce: the memory of the mirrors' new pieces, bound before this recording is submitted)
memory_table_flush(); // (vk_memory.cpp: the page table's changes, after this recording's mirror copies)
if (up_used) full_barrier(up_cmd); // (before the recording that reads them)
return up_used;
@@ -2315,6 +2341,13 @@ namespace {
bool upload_defer(const Texture& tex) {
static const int64_t budget_ns = (int64_t)(getenv("RX_VK_UPLOAD_BUDGET_MS") ? atof(getenv("RX_VK_UPLOAD_BUDGET_MS")) : 40.0) * 1000000;
if (!budget_ns || !upload_defer_ok || tex.uploaded || !tex.astc || tex.gpu_written) return false;
// (Triforce) and the first RX_VK_UPLOAD_MIN_ASTC of a frame (24) are never left for later. The budget counts every upload of
// the frame, and a frame whose other uploads took it all left every new ASTC texture on the placeholder: on a card at its
// VRAM allowance the render targets evicted and made again are uploaded from memory on the CPU, 58 to 190 ms a frame for
// frames on end (an RX 5700 XT's log, 2026-11-11: 8,586 textures left for later in one session), and the materials stayed
// flat white or blue for as long as that went on. An ASTC upload is decoded on the GPU: its part of such a frame is small
static const int min_astc = getenv("RX_VK_UPLOAD_MIN_ASTC") ? std::max(0, atoi(getenv("RX_VK_UPLOAD_MIN_ASTC"))) : 24;
if (frame_astc_admitted.fetch_add(1, std::memory_order_relaxed) < min_astc) return false;
if (frame_upload_ns.load(std::memory_order_relaxed) < budget_ns) return false;
st_deferred.fetch_add(1, std::memory_order_relaxed), st_deferred_frame.fetch_add(1, std::memory_order_relaxed);
return true;
@@ -2402,7 +2435,7 @@ Texture* texture_touch(Texture* tex, bool for_sampling) {
"%llu)", tex->id, (unsigned long long)tex->gpu, tex->width, tex->height, tex->layers, tex->nvn_format,
(unsigned long long)frame_no, (unsigned long long)tex->last_written);
}
if (tex->uploaded && !gpu_dirty && !texture_dirty(*tex, since, dirty)) return tex;
if (tex->uploaded && !gpu_dirty && !texture_dirty(*tex, since, dirty, !tex->gpu_written)) return tex;
const uint32_t gen = memory_generation();
// (storage render targets took last, which the CPU has not written since: nothing to upload, the memory holds
// nothing the GPU's images did not replace. The game's transient heaps move with the dynamic resolution, and a
@@ -2488,6 +2521,35 @@ Texture* texture_touch(Texture* tex, bool for_sampling) {
// (hookshot) a BC3 image whose last upload failed waits before the next try: its upload stages the whole ASTC source (some
// 200 MB for a terrain array), not to be repeated every submission while the memory it needs is not there
if (tex->bc && !tex->uploaded && tex->bc_retry_frame > frame_no) return tex;
// (Triforce, RX_VK_STAMP_SRC=1, a diagnostic: a texture of 1 MB or more uploaded again: who wrote its first dirty page (the
// write watch 1, a capture 2, else a caller of memory_stamp, placed by memory_stamp's own address), and whether its memory
// holds other bytes than at its last upload at all. 400 lines)
static const bool stamp_src = getenv("RX_VK_STAMP_SRC") != nullptr;
if (stamp_src && tex->uploaded && tex->guest_size >= (1u << 20)) {
static int told;
static std::unordered_map<uint32_t, int> times; // (RX_VK_STAMP_SRC=<n>: only a texture's n-th such upload and later)
static const int from_nth = std::max(1, atoi(getenv("RX_VK_STAMP_SRC")));
const uint64_t h = guest_hash(*tex);
if (++times[tex->id] >= from_nth && told++ < 400) {
uint64_t part = 0;
for (const auto& r : dirty) part += r.second - r.first;
const uint64_t off = dirty.empty() ? 0 : dirty[0].first;
const uint64_t guest = tex->gpu < (1ull << 40) ? tex->guest + off : host->gpu_guest(tex->gpu + off);
uint32_t pg = 0;
uintptr_t anchor = 0;
const uintptr_t who = guest ? memory_stamp_who(guest, &pg, &anchor) : 0;
logf("vk: stamp src: #%u %#llx %ux%u x%u NVN %#x uploaded again (%d) at frame %llu: %llu of %llu KB dirty in %zu ranges%s; bytes %s; "
"page %#llx gen %u (uploaded at %u) by %s%+lld", tex->id, (unsigned long long)tex->gpu, tex->width, tex->height,
tex->layers, tex->nvn_format, times[tex->id], (unsigned long long)frame_no, (unsigned long long)(part >> 10),
(unsigned long long)(tex->guest_size >> 10), dirty.size(), gpu_dirty ? " (a GPU buffer wrote it)" : "",
!tex->verify_hash ? "not hashed before" : h == tex->verify_hash ? "THE SAME as at its last upload" : "changed",
(unsigned long long)guest, pg, since, who == 1 ? "the write watch" : who == 2 ? "a capture" : who ? "memory_stamp" : "unknown",
who > 2 ? (long long)(who - anchor) : 0ll);
}
tex->verify_hash = h;
} else if (stamp_src && tex->guest_size >= (1u << 20)) {
tex->verify_hash = guest_hash(*tex);
}
const auto up0 = std::chrono::steady_clock::now();
const bool up_ok = upload(*tex, partial || sparse_written ? &dirty : nullptr);
{ // (an upload over 15 ms: what it was, 60 of them)
@@ -2565,6 +2627,84 @@ Texture* texture_touch(Texture* tex, bool for_sampling) {
return tex;
}
/// (Triforce, 2026-11-11) A flat texture sampled by a program whose instruction is a 3D sample: a one-slice 3D image with its
/// texels, and a 3D view of that. Vulkan has no 3D view of a 2D image, so such a binding had no view and read zeros (a null
/// descriptor). The console's GPU samples by the texture's own header whatever the instruction says: a flat texture's texel at
/// (u, v), the third coordinate unused, which a one-slice 3D image gives for any w. OoT's water in Zora's River binds one of
/// the game's 4 x 4 textures so (slot 0x102, 44 lookups in 10 s; its ripples ran in slow motion). Small plain textures only:
/// one layer copied, up to 1024 x 1024, not depth, not kept as BC3. Copied again at a lookup after the texture's contents
/// changed (an upload, or the GPU writing it): the lookups are cached by slot, so one the GPU draws every frame is a frame or
/// more behind here. RX_VK_NO_TWIN3D=1: no such view, as before
VkImageView twin3d_view(Texture* tex, VkFormat f, uint32_t level0, uint32_t nlevels, uint32_t layer, const VkComponentMapping& cm) {
static const bool off = getenv("RX_VK_NO_TWIN3D") != nullptr;
if (off || tex->type != VK_IMAGE_TYPE_2D || tex->sparse || tex->bc || tex->samples != VK_SAMPLE_COUNT_1_BIT ||
tex->aspect != VK_IMAGE_ASPECT_COLOR_BIT || is_bc(tex->format) || is_astc(tex->format) ||
(uint64_t)tex->width * tex->height > 1024ull * 1024 || !tex->image)
return VK_NULL_HANDLE;
VkCommandBuffer cmd = VK_NULL_HANDLE;
if (!tex->twin3d) {
VkImageCreateInfo ci{VK_STRUCTURE_TYPE_IMAGE_CREATE_INFO};
ci.imageType = VK_IMAGE_TYPE_3D;
ci.format = tex->format;
ci.extent = {tex->width, tex->height, 1};
ci.mipLevels = tex->levels;
ci.arrayLayers = 1;
ci.samples = VK_SAMPLE_COUNT_1_BIT;
ci.tiling = VK_IMAGE_TILING_OPTIMAL;
ci.usage = VK_IMAGE_USAGE_SAMPLED_BIT | VK_IMAGE_USAGE_TRANSFER_DST_BIT;
ci.flags = VK_IMAGE_CREATE_MUTABLE_FORMAT_BIT; // (the view's format may be another of its class: sRGB or not)
if (!create_image(ci, tex->twin3d, tex->twin3d_alloc, false, false, false)) {
tex->twin3d = VK_NULL_HANDLE, tex->twin3d_alloc = nullptr;
return VK_NULL_HANDLE;
}
cmd = upload_cmd();
VkImageMemoryBarrier2 b{VK_STRUCTURE_TYPE_IMAGE_MEMORY_BARRIER_2};
b.srcStageMask = b.dstStageMask = VK_PIPELINE_STAGE_2_ALL_COMMANDS_BIT;
b.dstAccessMask = VK_ACCESS_2_MEMORY_READ_BIT | VK_ACCESS_2_MEMORY_WRITE_BIT;
b.oldLayout = VK_IMAGE_LAYOUT_UNDEFINED, b.newLayout = VK_IMAGE_LAYOUT_GENERAL;
b.srcQueueFamilyIndex = b.dstQueueFamilyIndex = VK_QUEUE_FAMILY_IGNORED;
b.image = tex->twin3d;
b.subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, 0, VK_REMAINING_MIP_LEVELS, 0, 1};
VkDependencyInfo di{VK_STRUCTURE_TYPE_DEPENDENCY_INFO};
di.imageMemoryBarrierCount = 1, di.pImageMemoryBarriers = &b;
vkCmdPipelineBarrier2(cmd, &di);
tex->twin3d_value = ~0ull, tex->twin3d_written = ~0ull;
static int told;
if (told++ < 12)
logf("vk: texture #%u %#llx %ux%u levels %u NVN %#x is sampled as a 3D texture: a one-slice 3D copy of it made", tex->id,
(unsigned long long)tex->gpu, tex->width, tex->height, tex->levels, tex->nvn_format);
}
if (tex->twin3d_value != tex->upload_value || tex->twin3d_written != tex->last_written || tex->twin3d_layer != layer) {
if (!cmd) cmd = upload_cmd();
full_barrier(cmd, "barrier (a flat texture's 3D copy: before)");
std::vector<VkImageCopy> rs;
for (uint32_t l = 0; l < tex->levels; l++) {
VkImageCopy r{};
r.srcSubresource = {VK_IMAGE_ASPECT_COLOR_BIT, l, layer, 1};
r.dstSubresource = {VK_IMAGE_ASPECT_COLOR_BIT, l, 0, 1};
r.extent = {std::max(1u, tex->width >> l), std::max(1u, tex->height >> l), 1};
rs.push_back(r);
}
vkCmdCopyImage(cmd, tex->image, VK_IMAGE_LAYOUT_GENERAL, tex->twin3d, VK_IMAGE_LAYOUT_GENERAL, (uint32_t)rs.size(), rs.data());
full_barrier(cmd, "barrier (a flat texture's 3D copy: after)");
tex->twin3d_value = tex->upload_value, tex->twin3d_written = tex->last_written, tex->twin3d_layer = layer;
}
uint64_t key = mix((uint64_t)f, (uint64_t)level0 << 32 | nlevels);
key = mix(key, (uint64_t)cm.r << 24 | (uint64_t)cm.g << 16 | (uint64_t)cm.b << 8 | cm.a);
const auto it = tex->twin3d_views.find(key);
if (it != tex->twin3d_views.end()) return it->second;
VkImageViewCreateInfo vi{VK_STRUCTURE_TYPE_IMAGE_VIEW_CREATE_INFO};
vi.image = tex->twin3d;
vi.viewType = VK_IMAGE_VIEW_TYPE_3D;
vi.format = f;
vi.components = cm;
vi.subresourceRange = {VK_IMAGE_ASPECT_COLOR_BIT, level0, nlevels, 0, 1};
VkImageView view = VK_NULL_HANDLE;
if (vkCreateImageView(dev.device, &vi, nullptr, &view) != VK_SUCCESS) view = VK_NULL_HANDLE;
tex->twin3d_views[key] = view;
return view;
}
VkImageView texture_view(Texture* tex, const nvn_tex* t, const nvn_view* v, VkImageViewType type, ViewUse use,
VkFormat* format_out, VkFormat force_format) {
if (!tex) return VK_NULL_HANDLE;
@@ -2615,6 +2755,7 @@ VkImageView texture_view(Texture* tex, const nvn_tex* t, const nvn_view* v, VkIm
else if (type == VK_IMAGE_VIEW_TYPE_2D && nlayers > 1) nlayers = 1;
else if (type == VK_IMAGE_VIEW_TYPE_1D && nlayers > 1) nlayers = 1;
}
int twin_layer = -1; // (Triforce: a flat texture wanted as 3D: the layer its 3D copy is of)
// the shader's type over the image's: what Vulkan allows of them
switch (type) {
case VK_IMAGE_VIEW_TYPE_1D: case VK_IMAGE_VIEW_TYPE_1D_ARRAY:
@@ -2636,7 +2777,11 @@ VkImageView texture_view(Texture* tex, const nvn_tex* t, const nvn_view* v, VkIm
if (type == VK_IMAGE_VIEW_TYPE_2D) nlayers = 1;
break;
case VK_IMAGE_VIEW_TYPE_3D:
if (tex->type != VK_IMAGE_TYPE_3D) return VK_NULL_HANDLE;
if (tex->type != VK_IMAGE_TYPE_3D) {
// (Triforce: a flat texture sampled as a 3D one: its one-slice 3D copy, twin3d_view above)
if (use != ViewUse::Sample) return VK_NULL_HANDLE;
twin_layer = (int)layer0;
}
layer0 = 0, nlayers = 1;
break;
case VK_IMAGE_VIEW_TYPE_CUBE: case VK_IMAGE_VIEW_TYPE_CUBE_ARRAY:
@@ -2675,6 +2820,10 @@ VkImageView texture_view(Texture* tex, const nvn_tex* t, const nvn_view* v, VkIm
nf, (int)f, sw[0], sw[1], sw[2], sw[3], d.swizzle_r, d.swizzle_g, d.swizzle_b, d.swizzle_a,
v && v->set_swizzle ? 1 : 0);
}
if (twin_layer >= 0) {
if (format_out) *format_out = f;
return twin3d_view(tex, f, level0, nlevels, (uint32_t)twin_layer, cm);
}
uint64_t key = mix((uint64_t)use, (uint64_t)type);
key = mix(key, (uint64_t)f << 32 | aspect);
key = mix(key, (uint64_t)level0 << 48 | (uint64_t)nlevels << 32 | (uint64_t)layer0 << 16 | nlayers);
@@ -3948,6 +4097,7 @@ uint64_t preset_texture_budget(uint64_t preset_mb, bool preset_applies, bool vra
void texture_frame(uint64_t completed) {
upload_frame_report();
frame_upload_ns.store(0, std::memory_order_relaxed);
frame_astc_admitted.store(0, std::memory_order_relaxed);
if (const int64_t d = st_deferred_frame.exchange(0, std::memory_order_relaxed)) {
// (the views and descriptor sets resolved to the placeholder this frame go stale, as when textures go: the next frame
// looks those textures up again, and uploads them)
@@ -4129,6 +4279,7 @@ void texture_frame(uint64_t completed) {
}
if (budget < evict_to) evict_to = 0; // (hookshot, RX_VK_COMPACT: the driver's limit asks for more: 7/8 of that, as before)
if (frame_no % 600 == 0) {
memory_trial_tick(); // (Triforce: the memory handling on trial on this card, vk_memory.cpp)
vma_log_heaps();
vma_log_image_blocks();
{

View File

@@ -90,6 +90,13 @@ struct Texture {
uint64_t wb_value = 0; // (the submission whose own commands write its texels into its memory; 0: none)
uint64_t pool_key = 0; // (a render target's image: its create info's key in the pool it goes to when the texture does)
std::unordered_map<uint64_t, VkImageView> views;
// (Triforce) a flat texture a program samples as a 3D one: its one-slice 3D copy (vk_texture.cpp twin3d_view), the layer it
// was copied from and the texture's contents it was copied at
VkImage twin3d = VK_NULL_HANDLE;
VmaAllocation twin3d_alloc = nullptr;
uint32_t twin3d_layer = 0;
uint64_t twin3d_value = ~0ull, twin3d_written = ~0ull;
std::unordered_map<uint64_t, VkImageView> twin3d_views;
};
/// Makes texels [x, x + w) x [y, y + h) of level `level`, layers [first, first + count), of a sparse texture resident

View File

@@ -182,7 +182,40 @@ bool image_compact_pass_end(uint64_t* freed_bytes, uint32_t* freed_blocks) {
return compact_over(freed_bytes, freed_blocks);
}
bool create_image_inner(const VkImageCreateInfo& ci, VkImage& image, VmaAllocation& allocation, bool dedicated, bool target, bool textures);
/// (Triforce) An image that takes over 3 ms to make: what it was, how long, and whether its pool took a new block for it (60
/// lines). A player's session had frames of 50 to 150 ms in play that were all "texture lookups", 24 render targets made in one
bool create_image(const VkImageCreateInfo& ci, VkImage& image, VmaAllocation& allocation, bool dedicated, bool target, bool textures) {
const auto blocks = [&]() -> uint32_t {
VmaPool pool = textures ? texture_pool : target ? target_image_pool : image_pool;
if (!pool) return 0;
VmaStatistics st{};
vmaGetPoolStatistics(vma, pool, &st);
return st.blockCount;
};
const uint32_t before = blocks();
const auto t0 = std::chrono::steady_clock::now();
const bool ok = create_image_inner(ci, image, allocation, dedicated, target, textures);
const double ms = std::chrono::duration<double, std::milli>(std::chrono::steady_clock::now() - t0).count();
static std::atomic<int> told{0};
static std::atomic<int64_t> n{0}, slow{0};
static std::atomic<int64_t> slow_us{0};
n++;
if (ms > 3.0) {
slow++, slow_us += (int64_t)(ms * 1000);
if (told++ < 60) {
uint64_t used = 0, allowed = 0, block_free = 0;
vram_usage(&used, &allowed, &block_free);
logf("vk: an image took %.1f ms to make: %ux%u x%u levels %u format %d (%s)%s; its pool's blocks %u -> %u; VRAM %llu of %llu "
"MB, %llu MB free in the blocks; %lld of %lld images so far took over 3 ms (%.0f ms in all)", ms, ci.extent.width,
ci.extent.height, ci.arrayLayers, ci.mipLevels, (int)ci.format, textures ? "a texture" : target ? "a render target" : "other",
ok ? "" : ", NOT made", before, blocks(), (unsigned long long)(used >> 20), (unsigned long long)(allowed >> 20),
(unsigned long long)(block_free >> 20), (long long)slow.load(), (long long)n.load(), slow_us.load() / 1000.0);
}
}
return ok;
}
bool create_image_inner(const VkImageCreateInfo& ci, VkImage& image, VmaAllocation& allocation, bool dedicated, bool target, bool textures) {
VmaAllocationCreateInfo ai{};
ai.usage = VMA_MEMORY_USAGE_AUTO_PREFER_DEVICE;
// (in VRAM or not at all: an image VMA puts in system memory when VRAM is full is read over PCIe every frame,
@@ -278,6 +311,23 @@ bool create_device_buffer(uint64_t size, VkBufferUsageFlags usage, VkBuffer& buf
if (dedicated) ai.flags |= VMA_ALLOCATION_CREATE_DEDICATED_MEMORY_BIT;
return vmaCreateBuffer(vma, &bi, &ai, &buffer, &allocation, nullptr) == VK_SUCCESS;
}
/// (Triforce) Device memory for a piece of a sparse buffer (vk_memory.cpp's mirrors in pieces): size bytes at that alignment,
/// of a type in type_bits, within the driver's budget; the memory and the offset in it to bind
bool vma_piece_alloc(uint64_t size, uint64_t alignment, uint32_t type_bits, VmaAllocation& allocation, VkDeviceMemory& memory,
uint64_t& offset) {
VkMemoryRequirements req{};
req.size = size, req.alignment = alignment, req.memoryTypeBits = type_bits;
VmaAllocationCreateInfo ai{};
ai.requiredFlags = VK_MEMORY_PROPERTY_DEVICE_LOCAL_BIT;
if (dev.memory_budget) ai.flags = VMA_ALLOCATION_CREATE_WITHIN_BUDGET_BIT;
VmaAllocationInfo info{};
if (vmaAllocateMemory(vma, &req, &ai, &allocation, &info) != VK_SUCCESS) return false;
memory = info.deviceMemory, offset = info.offset;
return true;
}
void vma_piece_free(VmaAllocation allocation) {
vmaFreeMemory(vma, allocation);
}
/// (every heap: what VMA has in blocks and allocations there, what the process uses of it, its budget; one line)
/// The image pools: their blocks, the images in them, and the free ranges between them (every 600 presents, with the heaps)
void vma_log_image_blocks() {