Reflex: one present ID per frame from its start to its present; mode again on a new swapchain; presentId feature

The latency markers and the present carried different present IDs: g_reflex_present_id went up only in flip_locked, so
frame k's simulation, input and submits had ID k and its present k+1 (and with frame_queue or replay_async frames could
share one). Each frame now gets its ID once, when the game acquires its window texture (SIMULATION_START,
vk_reflex_frame_begin), and carries it: INPUT_SAMPLE and SIMULATION_END take the game's current frame, a submission
the frame of the thread that submits it (the game thread's, or the one the replay thread got with its ReplayItem, or
the present thread's from its PresentJob), and the present the frame of its blit (PresentJob.pid -> g_flip_pid). A
present without a frame ID takes the next value, as present IDs must grow on a swapchain.

From the review: a new swapchain sets the sleep mode again (it starts with low latency off), Reflex is off without the
presentId feature, REFLEX sorted in known[], and the ru and pt-BR descriptions name the Boost option as the menu does.

TOTK_TRACE=perf logs every 600 presents the frame start to present time, and with Reflex on how many of the driver's
reports (vkGetLatencyTimingsNV) came back with every marker and their simulation start to GPU render end.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
itaskaev2andClaude Opus 5.5 committed 2026-11-10 18:20:16 +01:00
1 parent 41623cec85
commit 6d6b66461b
9 files changed
+123 -31

No files matched your search

+86 -9
View File
@@ -266,7 +266,16 @@ static PFN_vkLatencySleepNV g_reflex_sleep;
static PFN_vkSetLatencyMarkerNV g_reflex_marker;
static VkSemaphore g_reflex_sem;
static uint64_t g_reflex_val;
static uint64_t g_reflex_present_id = 1;
// Present IDs (VK_KHR_present_id, VK_NV_low_latency2): each frame gets its ID once, when the game acquires its window
// texture (SIMULATION_START, vk_reflex_frame_begin), and the frame carries it to its input sample, its submissions
// (vk_reflex_submit_id: the game thread's current frame, or the frame a replay or present thread set for the work it
// does) and its present (PresentJob.pid, vk_reflex_present_frame, flip_locked), so the driver can match every marker
// of a frame to the present of that frame.
static uint64_t g_reflex_frame_next; // the last ID handed out
static uint64_t g_reflex_frame; // the frame the game simulates now
static _Thread_local uint64_t t_reflex_submit; // the frame whose work this thread submits; 0: g_reflex_frame
static uint64_t g_blit_pid, g_flip_pid; // the frame of the blit being made and of the present waiting (g_pl)
static uint64_t g_last_pid; // the last present ID presented (g_pl): present IDs only go up
static int g_reflex_mode = -1; // 0=off, 1=on, 2=boost
static uint32_t g_reflex_min_us = 1;
@@ -318,16 +327,75 @@ int vk_has_reflex(void) {
return c && c->has_nv_reflex;
}
uint64_t vk_reflex_current_present_id(void) {
return __atomic_load_n(&g_reflex_present_id, __ATOMIC_RELAXED);
// TOTK_TRACE=perf, every LAT_EVERY presents: the time from a frame's start (its window texture acquired) to its
// present call on the host, with Reflex off or on, and with Reflex on the driver's own reports of the same frames
// (vkGetLatencyTimingsNV): how many came back with every marker matched to their present, and their simulation start
// to GPU render end. The frame's start is kept by its ID, so it also shows that a present carries its frame's ID.
#define LAT_RING 64
#define LAT_EVERY 600
static struct { uint64_t id, start_ns; } g_lat_start[LAT_RING];
static struct { uint64_t n, ns, presents; } g_lat;
static PFN_vkGetLatencyTimingsNV g_reflex_timings;
uint64_t vk_reflex_frame_begin(void) {
uint64_t id = __atomic_add_fetch(&g_reflex_frame_next, 1, __ATOMIC_RELAXED);
__atomic_store_n(&g_reflex_frame, id, __ATOMIC_RELAXED);
if (rt_trace & RT_TRACE_PERF) g_lat_start[id % LAT_RING].start_ns = kern_now_ns(), g_lat_start[id % LAT_RING].id = id;
return id;
}
void vk_reflex_marker(int marker) {
// g_pl held, after the present of frame pid
static void latency_note(VkCtx *c, uint64_t pid) {
if (!(rt_trace & RT_TRACE_PERF) || !pid) return;
if (g_lat_start[pid % LAT_RING].id == pid) g_lat.n++, g_lat.ns += kern_now_ns() - g_lat_start[pid % LAT_RING].start_ns;
if (++g_lat.presents % LAT_EVERY) return;
char drv[200] = "";
if (g_reflex_mode > 0 && g_sc) {
if (!g_reflex_timings) g_reflex_timings = (PFN_vkGetLatencyTimingsNV)vkGetDeviceProcAddr(c->device, "vkGetLatencyTimingsNV");
static VkLatencyTimingsFrameReportNV rep[64];
VkGetLatencyMarkerInfoNV gi = {.sType = VK_STRUCTURE_TYPE_GET_LATENCY_MARKER_INFO_NV};
if (g_reflex_timings) {
g_reflex_timings(c->device, g_sc, &gi); // the count first
if (gi.timingCount > 64) gi.timingCount = 64;
for (uint32_t i = 0; i < gi.timingCount; i++)
rep[i] = (VkLatencyTimingsFrameReportNV){.sType = VK_STRUCTURE_TYPE_LATENCY_TIMINGS_FRAME_REPORT_NV};
gi.pTimings = rep;
if (gi.timingCount) g_reflex_timings(c->device, g_sc, &gi);
uint32_t full = 0;
double sum = 0;
for (uint32_t i = 0; i < gi.timingCount; i++) {
const VkLatencyTimingsFrameReportNV *r = &rep[i];
if (r->simStartTimeUs && r->simEndTimeUs && r->renderSubmitStartTimeUs && r->presentStartTimeUs &&
r->gpuRenderEndTimeUs > r->simStartTimeUs)
full++, sum += (double)(r->gpuRenderEndTimeUs - r->simStartTimeUs);
}
snprintf(drv, sizeof drv, "; driver reports: %u of %u frames with every marker, simulation start to GPU render "
"end %.2f ms", full, gi.timingCount, full ? sum / full / 1000.0 : 0.0);
}
}
rt_log("reflex: frame start to present %.2f ms over %llu frames (reflex %s)%s", g_lat.n ? g_lat.ns / 1e6 / g_lat.n : 0.0,
(unsigned long long)g_lat.n, g_reflex_mode == 2 ? "boost" : g_reflex_mode == 1 ? "on" : "off", drv);
g_lat.n = g_lat.ns = 0;
}
uint64_t vk_reflex_frame(void) { return __atomic_load_n(&g_reflex_frame, __ATOMIC_RELAXED); }
void vk_reflex_submit_frame(uint64_t id) { t_reflex_submit = id; }
uint64_t vk_reflex_submit_id(void) { return t_reflex_submit ? t_reflex_submit : vk_reflex_frame(); }
// The present thread, before the blit of frame id: the present of that blit carries id.
void vk_reflex_present_frame(uint64_t id) {
plat_rmutex_lock(&g_pl);
g_blit_pid = id;
plat_rmutex_unlock(&g_pl);
}
void vk_reflex_marker(int marker, uint64_t pid) {
VkCtx *c = vk_ctx();
if (!c || !c->has_nv_reflex || !g_reflex_marker || g_reflex_mode <= 0 || !g_sc) return;
if (!c || !c->has_nv_reflex || !g_reflex_marker || g_reflex_mode <= 0 || !g_sc || !pid) return;
plat_rmutex_lock(&g_pl);
if (g_sc && g_reflex_marker) {
uint64_t pid = __atomic_load_n(&g_reflex_present_id, __ATOMIC_RELAXED);
VkSetLatencyMarkerInfoNV mi = {
.sType = VK_STRUCTURE_TYPE_SET_LATENCY_MARKER_INFO_NV,
.presentID = pid,
@@ -400,11 +468,14 @@ static void flip_locked(void) {
uint64_t pid = 0;
VkPresentIdKHR pid_info = {0};
if (c && c->has_nv_reflex) {
pid = __atomic_add_fetch(&g_reflex_present_id, 1, __ATOMIC_RELAXED);
// The frame's own ID; a present without one (a probe, a skipped frame's neighbour) takes the next free value,
// as present IDs must grow on a swapchain
pid = g_flip_pid > g_last_pid ? g_flip_pid : g_last_pid + 1;
g_last_pid = pid;
pid_info.sType = VK_STRUCTURE_TYPE_PRESENT_ID_KHR;
pid_info.swapchainCount = 1;
pid_info.pPresentIds = &pid;
vk_reflex_marker(4); // VK_LATENCY_MARKER_PRESENT_START_NV
vk_reflex_marker(4, pid); // VK_LATENCY_MARKER_PRESENT_START_NV
}
VkPresentInfoKHR pi = {.sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR,
.pNext = (c && c->has_nv_reflex) ? &pid_info : NULL,
@@ -414,7 +485,8 @@ static void flip_locked(void) {
vk_queue_lock();
VkResult r = vkQueuePresentKHR(c->gfx_queue, &pi);
vk_queue_unlock();
if (c && c->has_nv_reflex) vk_reflex_marker(5); // VK_LATENCY_MARKER_PRESENT_END_NV
if (c && c->has_nv_reflex) vk_reflex_marker(5, pid); // VK_LATENCY_MARKER_PRESENT_END_NV
if (c) latency_note(c, pid ? pid : g_flip_pid);
vk_device_fault_report(r);
if (r == VK_ERROR_OUT_OF_DATE_KHR || r == VK_SUBOPTIMAL_KHR) plat_xchg32(&g_dirty, 1);
}
@@ -502,6 +574,10 @@ static int build_swapchain(void) {
}
free_capture();
vk_queue_unlock();
// The Reflex sleep mode belongs to the swapchain: a new one (resize, fullscreen, out of date) starts with low
// latency off, so the mode is set on it again even though the setting did not change.
g_reflex_mode = -1;
reflex_apply_locked();
return 1;
}
@@ -872,6 +948,7 @@ static void present_image(VkImage src, VkImageLayout layout, VkFormat fmt, int w
if (r == VK_SUCCESS) {
g_flip_on = 1;
g_flip_idx = idx;
g_flip_pid = g_blit_pid; // the frame this blit shows (vk_reflex_present_frame)
if (!defer) flip_locked();
} else {
vk_device_fault_report(r);