// (hookshot) NVIDIA Reflex through VK_NV_low_latency2, without NVIDIA's Streamline (vk_reflex.h). // // While the Insert menu's NVIDIA Reflex row (Off / On / On + Boost; RX_VK_REFLEX=0 / 1 / 2 over it) is on, the swapchain is // made with latency mode on (VkSwapchainLatencyCreateInfoNV) and its sleep mode set from the row; turning the row on or off // makes the swapchain again (reflex_swapchain_stale, at the next present). Off is the game as without Reflex: no latency // mode, no markers, no sleep, no present ids, nothing logged. The device's extensions are enabled whenever it has them, so // the row needs no restart: enabled alone they do nothing, nothing of them is used until the row is on. Each frame the driver gets markers with the // present's id: the game's simulation start (and input) once the sleep returns, its end and the rendering's submission start // at the game's present (do_present), the submission's end before the present's submit, and the present's start and end // around vkQueuePresentKHR. vkLatencySleepNV after the present waits until the driver wants the next frame to start: with // sleep mode off it returns at once. // // What the markers stand for is approximate: OoT simulates on its main thread and renders on its presentation thread, and the // replay records the GPU work as the game submits it; the sleep holds the presentation thread, which the main thread waits for. // With frame generation on, presents come from the display thread: Reflex is not used then. // // RX_VK_REFLEX_LOG=1: every 300 frames with Reflex on, the driver's own timings (vkGetLatencyTimingsNV) averaged in the log: // simulation start to present end, and to the GPU's end of rendering. // // (Triforce, after NX2navi's runtime/vk_lowlat.c, chrg6's: the driver holds the frame limit) RX_VK_REFLEX_LIMIT=1: with // Reflex on and no frame generation, its sleep also holds the Game frame rate cap (the sleep mode's minimum interval) and the // game's own wait (host/nn_vi.c hos_vi_native_acquire) steps aside, as the Reflex frame limit row does with DLSS frame // generation (vk_dlssg.cpp pace): one wait a frame, at its start, timed by the driver, which knows the GPU's time and the // last present. As there, the sleep that holds a frame has to be where the frame starts, so while Reflex holds the cap it is // taken on the game's thread at the acquire (reflex_pace, asked by the host through HOS_OPT_PACE_START: vk_pace.cpp) and // the one after the present is left out. Who holds a frame is dlssg_pace.h's rule with one frame shown for each (NX2navi's // vk_lowlat_pace): Reflex only while the driver has the very interval asked for, a present went out within 200 ms, no // movie plays, RivaTuner does not hold the game, and the frames have not shown that the sleep does not hold them (over half // of 60 sooner than three quarters of the period: the game's own wait for the rest of the run); the game's own wait in every // other case, from that frame on. Unset: no interval, the sleep after the present, the game's own wait, as before. #include "vk_reflex.h" #include #include #include #include #include #include "dlssg_pace.h" // ((Triforce) who holds the cap, Reflex or the game's own wait: reflex_pace) #include "vk_dlssg.h" #include "lab_ab.h" // (Triforce, lab: RX_VK_LAB_AB "noreflex") #include "vk_framegen.h" #include "vk_menu.h" #include "vk_work.h" // (last: RX_VK_RECORDER=defer drains before the latency-sleep wait; =thread does not need to, the driver signals it) namespace mkwvk { namespace { struct Reflex { bool have_ext = false, ready = false; bool latency_next = false; // (the swapchain being made has latency mode on: reflex_swapchain_info) bool latency = false; // (the current one has) VkPhysicalDevicePresentIdFeaturesKHR present_id{VK_STRUCTURE_TYPE_PHYSICAL_DEVICE_PRESENT_ID_FEATURES_KHR}; VkSwapchainLatencyCreateInfoNV latency_info{VK_STRUCTURE_TYPE_SWAPCHAIN_LATENCY_CREATE_INFO_NV}; PFN_vkSetLatencySleepModeNV set_mode = nullptr; PFN_vkLatencySleepNV sleep = nullptr; PFN_vkSetLatencyMarkerNV marker = nullptr; PFN_vkGetLatencyTimingsNV timings = nullptr; VkSwapchainKHR swapchain = VK_NULL_HANDLE; uint64_t swapchain_generation = 0; // (a sleep may finish after this chain was replaced) VkSemaphore wake = VK_NULL_HANDLE; // (timeline: the sleep's end) uint64_t wake_value = 0; uint64_t id = 1; // (the frame's present id) int mode = -1; // (as set on the swapchain) bool frame_marked = false; // (this frame's start markers given: its other markers count) // ((Triforce) the frame whose present comes next, as its submit ended (reflex_submit_ends): its present's id and markers. // With the present thread (vk_replay.cpp, RX_VK_PRESENT_THREAD) the present goes out after the game's thread has slept into // its next frame, whose id the sleep counted on) uint64_t shown_id = 0; bool shown_marked = false; VkPresentIdKHR pid{VK_STRUCTURE_TYPE_PRESENT_ID_KHR}; std::mutex mu; // ((Triforce) RX_VK_REFLEX_LIMIT: reflex_pace) uint32_t interval_us = 0; // (mu: the minimum interval as set on the swapchain with the mode) bool interval_refused = false; // (mu: the driver refused a mode with one: none is asked again) std::atomic want_us{0}; // (the interval the game's thread asked for, set at the next sleep) std::atomic game_sleeps{false}; // (the game's thread takes the sleep where its frame starts) std::atomic slept_after{false}; // (the sleep after a present was taken since the last frame's start) std::atomic hosted_ns{0}; // (the host's clock after the newest present with Reflex on, 0 before) DlssgPace pace{}; // (the game's thread's) int pace_said = -1; // (the game's thread's: who held the cap, as last logged) }; Reflex& rx() { static Reflex r; return r; } int wanted_mode() { static const char* e = getenv("RX_VK_REFLEX"); // (Triforce: with NVIDIA DLSS frame generation Reflex is Streamline's, vk_dlssg.cpp: this one stays off) if (dlssg::early()) return 0; return std::clamp(e && *e ? atoi(e) : opts.reflex, 0, 2); } /// With r.mu held: the row on, the extension and its feature, a swapchain made with latency mode, and no frame generation (its /// own display thread) bool on() { Reflex& r = rx(); return r.ready && r.swapchain && r.latency && wanted_mode() > 0 && !framegen_on(); } void mark(VkLatencyMarkerNV m) { Reflex& r = rx(); VkSetLatencyMarkerInfoNV i{VK_STRUCTURE_TYPE_SET_LATENCY_MARKER_INFO_NV}; i.presentID = r.id, i.marker = m; r.marker(dev.device, r.swapchain, &i); } /// ((Triforce) a marker of the frame whose present comes next: shown_id) void mark_shown(VkLatencyMarkerNV m) { Reflex& r = rx(); VkSetLatencyMarkerInfoNV i{VK_STRUCTURE_TYPE_SET_LATENCY_MARKER_INFO_NV}; i.presentID = r.shown_id, i.marker = m; r.marker(dev.device, r.swapchain, &i); } void apply_mode() { Reflex& r = rx(); const int want = wanted_mode(); // ((Triforce) and the frame limit the game's thread asked for, the minimum interval: reflex_pace; none unless asked) uint32_t us = want > 0 && !r.interval_refused ? r.want_us.load(std::memory_order_acquire) : 0; if (want == r.mode && us == r.interval_us) return; VkLatencySleepModeInfoNV m{VK_STRUCTURE_TYPE_LATENCY_SLEEP_MODE_INFO_NV}; m.lowLatencyMode = want >= 1, m.lowLatencyBoost = want == 2, m.minimumIntervalUs = us; VkResult res = r.set_mode(dev.device, r.swapchain, &m); if (res != VK_SUCCESS && us) { // ((Triforce) the interval refused: the mode as before, and the game's own wait holds the cap) r.interval_refused = true, us = 0, m.minimumIntervalUs = 0; logf("vk: NVIDIA Reflex: the driver refused a frame limit in the sleep mode (%d): the game's own wait holds the cap", (int)res); res = r.set_mode(dev.device, r.swapchain, &m); } if (want != r.mode) logf("vk: NVIDIA Reflex: %s%s", want == 0 ? "off" : want == 1 ? "on" : "on + boost", res == VK_SUCCESS ? "" : " (the driver refused the sleep mode)"); r.mode = want; r.interval_us = res == VK_SUCCESS ? us : 0; } void log_timings() { static const bool wanted = getenv("RX_VK_REFLEX_LOG") && atoi(getenv("RX_VK_REFLEX_LOG")) > 0; Reflex& r = rx(); if (!wanted || !r.timings || r.id % 300) return; VkGetLatencyMarkerInfoNV g{VK_STRUCTURE_TYPE_GET_LATENCY_MARKER_INFO_NV}; r.timings(dev.device, r.swapchain, &g); if (!g.timingCount) return; std::vector t(g.timingCount, {VK_STRUCTURE_TYPE_LATENCY_TIMINGS_FRAME_REPORT_NV}); g.pTimings = t.data(); r.timings(dev.device, r.swapchain, &g); double sim_present = 0, gpu_end = 0, sim = 0; int n = 0; for (uint32_t i = 0; i < g.timingCount; i++) { const auto& f = t[i]; if (!f.simStartTimeUs || !f.presentEndTimeUs || f.presentEndTimeUs < f.simStartTimeUs) continue; sim_present += (double)(f.presentEndTimeUs - f.simStartTimeUs); sim += f.simEndTimeUs > f.simStartTimeUs ? (double)(f.simEndTimeUs - f.simStartTimeUs) : 0.0; gpu_end += f.gpuRenderEndTimeUs > f.simStartTimeUs ? (double)(f.gpuRenderEndTimeUs - f.simStartTimeUs) : 0.0; n++; } if (n) logf("vk: NVIDIA Reflex (%s): over %d frames, simulation start to present end %.1f ms, to the GPU's end %.1f ms, " "simulation %.1f ms", r.mode == 0 ? "off" : r.mode == 1 ? "on" : "on + boost", n, sim_present / n / 1000.0, gpu_end / n / 1000.0, sim / n / 1000.0); } } // namespace void reflex_device_extensions(std::vector& want, const std::vector& all) { static const char* e = getenv("RX_VK_REFLEX"); if (e && !strcmp(e, "-")) return; // (RX_VK_REFLEX=-: the extensions not even enabled) bool ll = false, pid = false; for (const auto& x : all) { ll |= !strcmp(x.extensionName, VK_NV_LOW_LATENCY_2_EXTENSION_NAME); pid |= !strcmp(x.extensionName, VK_KHR_PRESENT_ID_EXTENSION_NAME); } if (!ll || !pid) return; for (const char* n : {VK_NV_LOW_LATENCY_2_EXTENSION_NAME, VK_KHR_PRESENT_ID_EXTENSION_NAME}) { bool listed = false; for (const char* w : want) listed |= !strcmp(w, n); if (!listed) want.push_back(n); } rx().have_ext = true; } void reflex_device_features(void**& tail) { Reflex& r = rx(); if (!r.have_ext) return; *tail = &r.present_id, tail = &r.present_id.pNext; } void reflex_device_created() { Reflex& r = rx(); if (!r.have_ext) return; if (!r.present_id.presentId) { logf("vk: NVIDIA Reflex: the device has no present ids: off"); return; } r.set_mode = (PFN_vkSetLatencySleepModeNV)vkGetDeviceProcAddr(dev.device, "vkSetLatencySleepModeNV"); r.sleep = (PFN_vkLatencySleepNV)vkGetDeviceProcAddr(dev.device, "vkLatencySleepNV"); r.marker = (PFN_vkSetLatencyMarkerNV)vkGetDeviceProcAddr(dev.device, "vkSetLatencyMarkerNV"); r.timings = (PFN_vkGetLatencyTimingsNV)vkGetDeviceProcAddr(dev.device, "vkGetLatencyTimingsNV"); VkSemaphoreTypeCreateInfo tc{VK_STRUCTURE_TYPE_SEMAPHORE_TYPE_CREATE_INFO}; tc.semaphoreType = VK_SEMAPHORE_TYPE_TIMELINE; VkSemaphoreCreateInfo sc{VK_STRUCTURE_TYPE_SEMAPHORE_CREATE_INFO}; sc.pNext = &tc; r.ready = r.set_mode && r.sleep && r.marker && vkCreateSemaphore(dev.device, &sc, nullptr, &r.wake) == VK_SUCCESS; logf("vk: NVIDIA Reflex: %s", r.ready ? "available (VK_NV_low_latency2)" : "its functions are missing: off"); } bool reflex_available() { return rx().ready; } void reflex_swapchain_info(VkSwapchainCreateInfoKHR& ci) { Reflex& r = rx(); r.latency_next = r.ready && wanted_mode() > 0; if (!r.latency_next) return; // (Off: the swapchain as without Reflex) r.latency_info.pNext = ci.pNext; r.latency_info.latencyModeEnable = VK_TRUE; ci.pNext = &r.latency_info; } bool reflex_swapchain_stale() { Reflex& r = rx(); std::lock_guard lk(r.mu); return r.ready && r.swapchain && r.latency != (wanted_mode() > 0); } void reflex_swapchain_created(VkSwapchainKHR swapchain) { Reflex& r = rx(); if (!r.ready) return; std::lock_guard lk(r.mu); const bool was = r.latency; ++r.swapchain_generation; r.swapchain = swapchain, r.mode = -1, r.frame_marked = false, r.shown_marked = false, r.latency = r.latency_next; r.interval_us = 0; // ((Triforce) a new swapchain has no frame limit yet: apply_mode sets the one asked for) if (r.latency) apply_mode(); else if (was) logf("vk: NVIDIA Reflex: off"); } void reflex_swapchain_destroyed() { Reflex& r = rx(); std::lock_guard lk(r.mu); ++r.swapchain_generation; r.swapchain = VK_NULL_HANDLE; r.latency = false; r.frame_marked = false, r.shown_marked = false; r.mode = -1; } void reflex_present_begins() { Reflex& r = rx(); std::lock_guard lk(r.mu); if (!on() || !r.frame_marked) return; mark(VK_LATENCY_MARKER_SIMULATION_END_NV); mark(VK_LATENCY_MARKER_RENDERSUBMIT_START_NV); } void reflex_submit_ends() { Reflex& r = rx(); std::lock_guard lk(r.mu); r.shown_marked = false; if (!on()) return; r.shown_id = r.id, r.shown_marked = r.frame_marked; // ((Triforce) the present's, whichever thread gives it) if (r.frame_marked) mark(VK_LATENCY_MARKER_RENDERSUBMIT_END_NV); } const void* reflex_present_id(const void* next) { Reflex& r = rx(); std::lock_guard lk(r.mu); if (!on() || !r.shown_marked) return next; mark_shown(VK_LATENCY_MARKER_PRESENT_START_NV); r.pid.pNext = next, r.pid.swapchainCount = 1, r.pid.pPresentIds = &r.shown_id; return &r.pid; } void reflex_presented() { Reflex& r = rx(); std::lock_guard lk(r.mu); if (!on()) return; if (r.shown_marked) mark_shown(VK_LATENCY_MARKER_PRESENT_END_NV); r.hosted_ns.store(host->now_ns(), std::memory_order_release); // ((Triforce) for reflex_pace: a present went out lately) } namespace { /// The driver's sleep until the next frame should start, then that frame's start markers: after the present /// (reflex_sleep), or where the frame starts while Reflex holds the cap (reflex_pace). False when no sleep was taken bool sleep_now() { Reflex& r = rx(); uint64_t generation, value; { std::lock_guard lk(r.mu); if (!on()) return false; generation = r.swapchain_generation; apply_mode(); // (the menu's row may have changed) if (r.frame_marked) log_timings(), r.id++; VkLatencySleepInfoNV s{VK_STRUCTURE_TYPE_LATENCY_SLEEP_INFO_NV}; s.signalSemaphore = r.wake, s.value = ++r.wake_value; value = s.value; if (r.sleep(dev.device, r.swapchain, &s) != VK_SUCCESS) return false; } VkSemaphoreWaitInfo w{VK_STRUCTURE_TYPE_SEMAPHORE_WAIT_INFO}; w.semaphoreCount = 1, w.pSemaphores = &r.wake, w.pValues = &value; vkWaitSemaphores(dev.device, &w, 100000000ull); // (100 ms at most: a driver that never wakes it does not stop the game) std::lock_guard lk(r.mu); if (!on() || generation != r.swapchain_generation || value != r.wake_value) return true; mark(VK_LATENCY_MARKER_SIMULATION_START_NV); mark(VK_LATENCY_MARKER_INPUT_SAMPLE_NV); r.frame_marked = true; return true; } /// (Triforce) Who holds the cap, as a log line when it or the reason changes (the first 40). why: the reason the game's own /// wait holds it void pace_say(int reflex, uint32_t us, const char* why) { Reflex& r = rx(); static int lines; static const char* said_why; if ((r.pace_said == reflex && (reflex || why == said_why)) || lines >= 40) return; r.pace_said = reflex, said_why = why, lines++; if (reflex) logf("vk: NVIDIA Reflex: frame limit: its sleep holds the frame rate cap where each frame starts (%.3f ms, " "RX_VK_REFLEX_LIMIT); the game's own wait steps aside", us / 1000.0); else logf("vk: NVIDIA Reflex: frame limit: the game's own wait holds it (%s)", why); } } // namespace void reflex_sleep() { // ((Triforce) while Reflex holds the Game frame rate cap the game's thread takes the sleep where its frame starts; a sleep // taken here is noted for the frame that starts next, reflex_pace) Reflex& r = rx(); if (r.game_sleeps.load(std::memory_order_acquire)) return; if (lab_ab::on("noreflex")) return; // (lab: the sleep after the present left out in these phases: what it costs the frame) if (sleep_now()) r.slept_after.store(true, std::memory_order_release); } bool reflex_pace(int64_t period_ns) { static const bool wanted = [] { const char* e = getenv("RX_VK_REFLEX_LIMIT"); return e && atoi(e) != 0; }(); if (!wanted) return false; Reflex& r = rx(); bool live, refused; uint32_t applied; { std::lock_guard lk(r.mu); live = on(), refused = r.interval_refused; applied = r.mode > 0 ? r.interval_us : 0; } const int64_t hosted = r.hosted_ns.load(std::memory_order_acquire), now = host->now_ns(); // (in this order: now is not before hosted) uint32_t want = 0; int found = DLSSG_PACE_SAME; // (dlssg_pace.h's rule with one frame shown for each of the game's: Reflex while the driver has the very interval asked // for, a present went out lately and its frames have not shown that it does not hold them) bool held = dlssg_pace_step(&r.pace, (uint64_t)now, period_ns > 0 ? (uint64_t)period_ns : 0, 1, (uint64_t)std::max(hosted, 0), live && !refused, applied, &want, &found) != 0; r.want_us.store(want, std::memory_order_release); // (set with the sleep mode at the next sleep: apply_mode) if (found == DLSSG_PACE_FAST) logf("vk: NVIDIA Reflex: its sleep does not hold the frame rate cap (over half of %d frames came sooner than three " "quarters of %.2f ms): the game's own wait holds it for the rest of this run", DLSSG_PACE_WINDOW, period_ns / 1e6); // (the frame's sleep here, where it starts, in the game's own wait's place; but where the sleep after the present before // it was taken, which is so for the first frame Reflex holds, that one, with the interval set, held this frame) const bool after = r.slept_after.exchange(false, std::memory_order_acq_rel); if (held && !after && !sleep_now()) held = false; // (none taken: the game's own wait) r.game_sleeps.store(held, std::memory_order_release); if (held) pace_say(1, want, nullptr); else if (!want) // (an interval the driver has yet to take is no change: it is set at the next sleep) pace_say(0, 0, !live ? "Reflex is off, or frame generation presents the frames" : refused ? "the driver refused the frame limit" : r.pace.bad ? "Reflex's sleep did not hold the frames" : period_ns <= 0 ? "a movie plays, or RivaTuner holds the game" : hosted ? "no present for 200 ms" : "no present yet"); return held; } } // namespace mkwvk