MacKwnoX/NX2navi
1
0
Fork 0
mirror of https://github.com/tangerineturtle1/NX2navi.git synced 2026-11-15 19:19:58 -04:00
Code Issues Packages Projects Releases Wiki Activity
Files
main
NX2navi/runtime/dlssg/nx2_sl.cpp
chrg6 1c3fa7df2a Give Streamline constants and null tags on a frame without inputs while frame generation is on. 
A review of the DLSS and frame generation paths against NVIDIA's guides (the DLSS programming guide 310.6.0, the
DLSS-FG guide 310.7.0, Streamline 2.14.16's guides and headers).

What follows them: the texture bias (log2(render width / output width) - 1 on top of the game's own, at each change
of the mode), the jitter (Halton, 8 x (output / render)^2 phases and 16 at least, in render pixels within half a
pixel, the scene drawn with it and the motion vectors made without it), the creation flags (HDR, vectors at the
render size, auto-exposure, depth not inverted) and the constants Streamline gets (unjittered matrices, the jitter
in pixels, the vectors' scale of one over the render size).

What did not: vk_dlssg.c holds a change of generation's state for FG_SETTLE_FRAMES frames, so for up to 12 frames
after the depth, the motion vectors or the camera went away (a menu, a fade, a load) generation stayed on and
nx2sl_frame set no constants and tagged nothing. Streamline's guide asks for the constants at every frame and for
null tags where depth or motion vectors are not valid; it logged "Unable to find 'common' constants for frame" and
"Failed to find global tag 'kBufferTypeDepth'" and used the frame before's (in every play log with frame
generation). Such a frame now gets the last frame's constants with reset set and the three tags as null. Not run in
a game yet: nx2navi_dlssg.dll builds and exports its 15 functions.

What the guides leave open: the sign of the jitter offsets against the shift of the picture, which NVIDIA says to
find by trying, and the bias reaches only the tiled, trilinear samplers, not every texture of the scene. For both
there is a test switch now, read once: NX2NAVI_DLAA_JITTER_TEST=<x>,<y> tells the upscaler the jitter times two
factors, and NX2NAVI_DLSS_LOD_BIAS_TEST=<bias> puts a bias in the rule's place. Neither was measured: the runs
that would do it start the game.

The line about the presents a second (the commit before) came at 59 a second under a limit of 60: it comes under 56
now.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
2026-11-14 12:54:08 -05:00

336 lines
18 KiB
C++
Raw Permalink Blame History

// nx2navi_dlssg.dll: DLSS Frame Generation through NVIDIA Streamline, behind the C interface of nx2_sl.h. Built with
// MSVC against the Streamline SDK's headers (tools/build_dlssg.py; the SDK stays outside the repository), so its C++
// types have the ABI of Streamline's own DLLs. sl.interposer.dll and its plugins are loaded from the game's folder at
// run time; this DLL links none of them.
//
// Manual hooking (the SDK's ProgrammingGuideManualHooking): the game creates its instance and device through the
// interposer's vkCreateInstance and vkCreateDevice, which add what DLSS-G needs, and calls the interposer's
// swapchain functions; everything else goes to the Vulkan loader as before. Each frame of the game sleeps in Reflex
// where it starts and gets the PCL latency markers, the camera constants and the depth and motion vector tags (valid
// until the present), then the present goes through the interposer, which inserts the generated frames: one, two or
// three between two rendered frames (2x, 3x or 4x), up to what the GPU and driver allow (numFramesToGenerateMax).
#define NX2SL_BUILD
#include <windows.h>
#include <stdio.h>
#include <string.h>
#include <atomic>
#include <sl.h>
#include <sl_consts.h>
#include <sl_dlss_g.h>
#include <sl_helpers.h>
#include <sl_pcl.h>
#include <sl_reflex.h>
#include "nx2_sl.h"
namespace {
HMODULE g_mod;
Nx2SlLog g_log;
PFun_slInit *p_init;
PFun_slShutdown *p_shutdown;
PFun_slIsFeatureLoaded *p_loaded;
PFun_slSetTagForFrame *p_tag;
PFun_slSetConstants *p_constants;
PFun_slGetNewFrameToken *p_token;
PFun_slGetFeatureFunction *p_feature;
PFun_slDLSSGSetOptions *p_dlssg_options;
PFun_slDLSSGGetState *p_dlssg_state;
PFun_slPCLSetMarker *p_marker;
PFun_slReflexSetOptions *p_reflex;
PFun_slReflexSleep *p_sleep;
PFun_slReflexGetState *p_reflex_state;
PFN_vkGetInstanceProcAddr p_gipa;
PFN_vkGetDeviceProcAddr p_gdpa;
int g_available = -1;
int g_low_latency = -1; // Reflex's low-latency sleep on this GPU and driver (ReflexState), -1 not asked
int g_frames = -1; // the frames to generate as last set (0 off), -1 before the first nx2sl_frame
std::atomic<uint32_t> g_max{0}; // numFramesToGenerateMax as Streamline last reported it, 0 before it did
uint64_t g_presented, g_shown;
void say(const char *fmt, ...) {
if (!g_log) return;
char s[512];
va_list a;
va_start(a, fmt);
vsnprintf(s, sizeof s, fmt, a);
va_end(a);
g_log(s);
}
void sl_log(sl::LogType type, const char *msg) {
if (type == sl::LogType::eInfo) return; // only warnings and errors reach the game's log
static int n;
if (n++ < 200) say("dlssg: streamline %s: %s", type == sl::LogType::eError ? "error" : "warning", msg);
}
template <class T> T *proc(const char *name) { return reinterpret_cast<T *>(GetProcAddress(g_mod, name)); }
// Whether Reflex's sleep runs on this GPU and driver: Streamline takes a frame limit without it and then sets nothing
// (its slReflexSetOptions hands the limit to the driver only with the low-latency sleep available). Asked until
// Streamline answers, from nx2sl_available and nx2sl_reflex2 (nx2sl_reflex too, which an older runtime calls), which
// the runtime calls on the device's thread before any present or under its present lock (vk_dlssg.c reflex_apply),
// never two at once: slReflexGetState is not thread safe.
void reflex_state() {
static sl::ReflexState rs;
if (g_low_latency < 0 && p_reflex_state && p_reflex_state(rs) == sl::Result::eOk)
g_low_latency = rs.lowLatencyAvailable ? 1 : 0;
}
// The token of frame n. Tokens come from a small ring Streamline reuses, so each call takes its own right before use.
sl::FrameToken *token(uint32_t frame) {
sl::FrameToken *t = nullptr;
return frame && p_token(t, &frame) == sl::Result::eOk ? t : nullptr;
}
void matrix(sl::float4x4 &m, const float *v) {
for (int r = 0; r < 4; r++) m[r] = sl::float4(v[r * 4], v[r * 4 + 1], v[r * 4 + 2], v[r * 4 + 3]);
}
sl::Resource resource(const Nx2SlImage *i) {
sl::Resource r(sl::ResourceType::eTex2d, (void *)i->image, nullptr, (void *)i->view, (uint32_t)i->layout);
r.width = i->width;
r.height = i->height;
r.nativeFormat = (uint32_t)i->format;
r.mipLevels = 1;
r.arrayLayers = 1;
r.usage = (uint32_t)i->usage;
return r;
}
} // namespace
extern "C" NX2SL_API int nx2sl_version(void) { return NX2SL_VERSION; }
extern "C" NX2SL_API int nx2sl_init(const wchar_t *dir, Nx2SlLog log) {
g_log = log;
wchar_t path[MAX_PATH];
swprintf(path, MAX_PATH, L"%ls\\sl.interposer.dll", dir);
g_mod = LoadLibraryW(path);
if (!g_mod) return say("dlssg: sl.interposer.dll not found next to the game: DLSS Frame Generation off"), 0;
p_init = proc<PFun_slInit>("slInit");
p_shutdown = proc<PFun_slShutdown>("slShutdown");
p_loaded = proc<PFun_slIsFeatureLoaded>("slIsFeatureLoaded");
p_tag = proc<PFun_slSetTagForFrame>("slSetTagForFrame");
p_constants = proc<PFun_slSetConstants>("slSetConstants");
p_token = proc<PFun_slGetNewFrameToken>("slGetNewFrameToken");
p_feature = proc<PFun_slGetFeatureFunction>("slGetFeatureFunction");
p_gipa = reinterpret_cast<PFN_vkGetInstanceProcAddr>(GetProcAddress(g_mod, "vkGetInstanceProcAddr"));
p_gdpa = reinterpret_cast<PFN_vkGetDeviceProcAddr>(GetProcAddress(g_mod, "vkGetDeviceProcAddr"));
if (!p_init || !p_tag || !p_constants || !p_token || !p_feature || !p_gipa || !p_gdpa)
return say("dlssg: sl.interposer.dll lacks its entry points"), 0;
static const sl::Feature features[] = {sl::kFeatureDLSS_G, sl::kFeatureReflex, sl::kFeaturePCL};
const wchar_t *paths[] = {dir};
sl::Preferences p{};
p.showConsole = false;
p.logLevel = sl::LogLevel::eDefault;
p.pathsToPlugins = paths;
p.numPathsToPlugins = 1;
p.pathToLogsAndData = dir;
p.logMessageCallback = sl_log;
// the game tracks its own command buffer state; no plugin downloads at run time. The depth and motion vectors are
// tagged per frame (slSetTagForFrame, nx2sl_frame), which Streamline refuses without the frame-based flag.
p.flags = sl::PreferenceFlags::eDisableCLStateTracking | sl::PreferenceFlags::eUseManualHooking |
sl::PreferenceFlags::eUseFrameBasedResourceTagging;
p.featuresToLoad = features;
p.numFeaturesToLoad = 3;
p.applicationId = 231313132; // the id NVIDIA's samples use for builds without an issued one (as vk_dlaa.c's NGX)
p.engine = sl::EngineType::eCustom;
p.renderAPI = sl::RenderAPI::eVulkan;
sl::Result r = p_init(p, sl::kSDKVersion);
if (r != sl::Result::eOk) return say("dlssg: slInit failed (%d): DLSS Frame Generation off", (int)r), 0;
say("dlssg: Streamline up");
return 1;
}
extern "C" NX2SL_API PFN_vkGetInstanceProcAddr nx2sl_instance_proc(void) { return p_gipa; }
extern "C" NX2SL_API PFN_vkGetDeviceProcAddr nx2sl_device_proc(void) { return p_gdpa; }
extern "C" NX2SL_API int nx2sl_available(void) {
if (g_available >= 0) return g_available;
bool loaded = false;
g_available = p_loaded && p_loaded(sl::kFeatureDLSS_G, loaded) == sl::Result::eOk && loaded;
void *f = nullptr;
if (g_available && p_feature(sl::kFeatureDLSS_G, "slDLSSGSetOptions", f) == sl::Result::eOk) p_dlssg_options = (PFun_slDLSSGSetOptions *)f;
if (g_available && p_feature(sl::kFeatureDLSS_G, "slDLSSGGetState", f) == sl::Result::eOk) p_dlssg_state = (PFun_slDLSSGGetState *)f;
if (p_feature(sl::kFeaturePCL, "slPCLSetMarker", f) == sl::Result::eOk) p_marker = (PFun_slPCLSetMarker *)f;
if (p_feature(sl::kFeatureReflex, "slReflexSetOptions", f) == sl::Result::eOk) p_reflex = (PFun_slReflexSetOptions *)f;
if (p_feature(sl::kFeatureReflex, "slReflexSleep", f) == sl::Result::eOk) p_sleep = (PFun_slReflexSleep *)f;
if (p_feature(sl::kFeatureReflex, "slReflexGetState", f) == sl::Result::eOk) p_reflex_state = (PFun_slReflexGetState *)f;
reflex_state();
if (!p_dlssg_options || !p_marker) g_available = 0;
// How many frames this GPU and driver generate per rendered frame: known once the device exists (else after the
// first presents, nx2sl_present_marker)
sl::DLSSGState s{};
if (g_available && p_dlssg_state && p_dlssg_state(sl::ViewportHandle(0), s, nullptr) == sl::Result::eOk)
g_max.store(s.numFramesToGenerateMax);
if (!g_available) say("dlssg: frame generation unavailable (DLSS-G plugin loaded: %d, DLSS-G options: %d, PCL marker: %d)",
loaded, p_dlssg_options != nullptr, p_marker != nullptr);
else if (g_max.load()) say("dlssg: DLSS Frame Generation available, up to %ux", g_max.load() + 1);
else say("dlssg: DLSS Frame Generation available (its most frames per rendered frame not reported yet)");
return g_available;
}
extern "C" NX2SL_API void nx2sl_log_unavailable(VkPhysicalDevice physical) {
auto supported = proc<PFun_slIsFeatureSupported>("slIsFeatureSupported");
if (!supported) return say("dlssg: Streamline has no adapter support query");
sl::AdapterInfo adapter{};
adapter.vkPhysicalDevice = physical;
sl::Result r = supported(sl::kFeatureDLSS_G, adapter);
say("dlssg: Streamline's support check for frame generation: %s (%d)", sl::getResultAsStr(r), (int)r);
}
// Streamline sets Reflex's mode and frame limit on the swapchain it was told of last (VK_NV_low_latency2, the sleep
// mode's minimum interval) and drops them while there is none, so both are set again after each swapchain. Every call
// carries both: a call without the limit clears it. 1 when Streamline took them, unless Streamline said Reflex's sleep
// does not run here (then nothing would hold the limit). Where it has not said, the runtime's own check of the frames
// (dlssg_pace.h) finds a limit nothing holds.
extern "C" NX2SL_API int nx2sl_reflex2(int mode, uint32_t frame_limit_us) {
if (g_available != 1 || !p_reflex) return 0;
reflex_state();
sl::ReflexOptions o{};
o.mode = mode == 2 ? sl::ReflexMode::eLowLatencyWithBoost : mode ? sl::ReflexMode::eLowLatency : sl::ReflexMode::eOff;
o.frameLimitUs = frame_limit_us;
sl::Result r = p_reflex(o);
static int n;
if (n++ < 8 || (r != sl::Result::eOk && n < 40)) { // the first calls, and later ones Streamline refused
char limit[48] = "no frame limit";
if (frame_limit_us) snprintf(limit, sizeof limit, "frame limit %.3f ms", frame_limit_us / 1000.0);
say("dlssg: Reflex low-latency mode %s, %s (%d)", mode == 2 ? "on with boost" : mode ? "on" : "off", limit, (int)r);
}
return r == sl::Result::eOk && (!frame_limit_us || g_low_latency != 0);
}
// The runtime before nx2sl_reflex2: the mode alone, no frame limit.
extern "C" NX2SL_API void nx2sl_reflex(int mode) { nx2sl_reflex2(mode, 0); }
// The game's simulation and its recording of the frame's commands are one span here: the guest's threads cannot be
// told apart, so the simulation markers sit at the frame's start and the render submission runs until the frame's
// submissions went out (nx2sl_submit_end).
extern "C" NX2SL_API void nx2sl_sleep(uint32_t frame, int hold) {
if (g_available != 1) return;
sl::FrameToken *t = token(frame);
if (hold && t && p_sleep) p_sleep(*t);
if ((t = token(frame))) {
p_marker(sl::PCLMarker::eSimulationStart, *t);
p_marker(sl::PCLMarker::eSimulationEnd, *t);
p_marker(sl::PCLMarker::eRenderSubmitStart, *t);
}
}
extern "C" NX2SL_API void nx2sl_submit_end(uint32_t frame) {
sl::FrameToken *t = g_available == 1 ? token(frame) : nullptr;
if (t) p_marker(sl::PCLMarker::eRenderSubmitEnd, *t);
}
extern "C" NX2SL_API int nx2sl_frame(VkCommandBuffer cb, const Nx2SlImage *depth, const Nx2SlImage *mv,
const Nx2SlImage *hudless, const Nx2SlCamera *cam, uint32_t out_w, uint32_t out_h,
int enable, uint32_t frame) {
if (g_available != 1) return 0;
sl::FrameToken *t = token(frame);
if (!t) return 0;
sl::ViewportHandle vp(0);
// never more than Streamline allows: one while it has not said
uint32_t max = g_max.load();
int n = enable > 0 ? enable : 0;
if (n > (int)(max ? max : 1)) n = (int)(max ? max : 1);
static sl::Constants last; // the constants of the last frame that had its inputs
static bool have_last;
if (n && !(depth && mv && cam) && have_last) {
// Generation is on for this present and the frame has no depth, motion vectors or camera: a menu, a fade or a
// load, while vk_dlssg.c holds the change of state for a few frames. Streamline's guide asks for the common
// constants at every frame and for null tags wherever depth or motion vectors are not valid
// (ProgrammingGuideDLSS_G.md 5.2 and 7.0). This frame gave neither, and Streamline logged "Unable to find
// 'common' constants for frame" and "Failed to find global tag 'kBufferTypeDepth'" and went on with what the
// frame before had set. It gets the last frame's constants again, as a new start, and the three tags as null.
sl::Constants c = last;
c.reset = sl::Boolean::eTrue;
p_constants(c, *t, vp);
sl::ResourceTag none[3] = {sl::ResourceTag(nullptr, sl::kBufferTypeDepth, sl::ResourceLifecycle::eValidUntilPresent),
sl::ResourceTag(nullptr, sl::kBufferTypeMotionVectors, sl::ResourceLifecycle::eValidUntilPresent),
sl::ResourceTag(nullptr, sl::kBufferTypeHUDLessColor, sl::ResourceLifecycle::eValidUntilPresent)};
if ((t = token(frame))) p_tag(*t, vp, none, 3, (sl::CommandBuffer *)cb);
static int said;
if (!said++) say("dlssg: a frame without depth and motion vectors while generation is on gets the last constants "
"as a new start and null tags");
}
if (n && depth && mv && cam) {
sl::Constants c{};
matrix(c.cameraViewToClip, cam->view_to_clip);
matrix(c.clipToCameraView, cam->clip_to_view);
matrix(c.clipToPrevClip, cam->clip_to_prev_clip);
matrix(c.prevClipToClip, cam->prev_clip_to_clip);
c.jitterOffset = sl::float2(cam->jitter[0], cam->jitter[1]);
c.mvecScale = sl::float2(cam->mv_scale[0], cam->mv_scale[1]);
c.cameraPinholeOffset = sl::float2(0, 0);
c.cameraPos = sl::float3(cam->pos[0], cam->pos[1], cam->pos[2]);
c.cameraUp = sl::float3(cam->up[0], cam->up[1], cam->up[2]);
c.cameraRight = sl::float3(cam->right[0], cam->right[1], cam->right[2]);
c.cameraFwd = sl::float3(cam->fwd[0], cam->fwd[1], cam->fwd[2]);
c.cameraNear = cam->znear;
c.cameraFar = cam->zfar;
c.cameraFOV = cam->fov_y;
c.cameraAspectRatio = cam->aspect;
c.depthInverted = sl::Boolean::eFalse;
c.cameraMotionIncluded = sl::Boolean::eTrue;
c.motionVectors3D = sl::Boolean::eFalse;
c.reset = cam->reset ? sl::Boolean::eTrue : sl::Boolean::eFalse;
p_constants(c, *t, vp);
last = c, have_last = true;
sl::Resource rd = resource(depth), rm = resource(mv), rh = hudless ? resource(hudless) : sl::Resource{};
sl::Extent ed{0, 0, depth->width, depth->height}, em{0, 0, mv->width, mv->height};
sl::Extent eh{0, 0, hudless ? hudless->width : 0, hudless ? hudless->height : 0};
sl::ResourceTag tags[3] = {sl::ResourceTag(&rd, sl::kBufferTypeDepth, sl::ResourceLifecycle::eValidUntilPresent, &ed),
sl::ResourceTag(&rm, sl::kBufferTypeMotionVectors, sl::ResourceLifecycle::eValidUntilPresent, &em),
sl::ResourceTag(&rh, sl::kBufferTypeHUDLessColor, sl::ResourceLifecycle::eValidUntilPresent, &eh)};
if ((t = token(frame))) p_tag(*t, vp, tags, hudless ? 3 : 2, (sl::CommandBuffer *)cb);
}
if (n != g_frames) {
sl::DLSSGOptions o{};
o.mode = n ? sl::DLSSGMode::eOn : sl::DLSSGMode::eOff;
o.numFramesToGenerate = n ? (uint32_t)n : 1;
// Generation goes off for every menu, cutscene and frame without depth and motion, and on again after it. Off
// frees the generator's resources unless this flag is set, and on allocates them again inside the present, with
// the queue lock held: a frame of 100 to 150 ms each time, the submitting thread waiting for that lock. With
// the flag, off only suspends. Whether frame generation runs at all is fixed for the run (vk_dlssg.c), so
// nothing is held for long that would otherwise be given back; Streamline frees it at shutdown.
o.flags = sl::DLSSGFlags::eRetainResourcesWhenOff;
sl::Result r = p_dlssg_options(vp, o);
if (n) say("dlssg: frame generation on, %dx (%d generated per rendered frame) (%d), output %ux%u", n + 1, n, (int)r, out_w, out_h);
else say("dlssg: frame generation off (%d), output %ux%u", (int)r, out_w, out_h);
g_frames = n;
}
return 1;
}
extern "C" NX2SL_API void nx2sl_present_marker(int after, uint32_t frame) {
sl::FrameToken *t = g_available == 1 ? token(frame) : nullptr;
if (!t) return;
p_marker(after ? sl::PCLMarker::ePresentEnd : sl::PCLMarker::ePresentStart, *t);
if (!after) return;
g_presented++;
// Asked after each present, numFramesActuallyPresented is what that present showed: 2 with one generated frame, 4
// with three.
sl::DLSSGState s{};
if (!p_dlssg_state || p_dlssg_state(sl::ViewportHandle(0), s, nullptr) != sl::Result::eOk) return;
g_shown += s.numFramesActuallyPresented;
if (s.numFramesToGenerateMax && s.numFramesToGenerateMax != g_max.load()) { // a 0 now and then is no news
say("dlssg: up to %ux on this GPU and driver", s.numFramesToGenerateMax + 1);
g_max.store(s.numFramesToGenerateMax);
}
if (g_presented == 600 || g_presented % 3600 == 0)
say("dlssg: %llu frames, %u presented per frame, status %u", (unsigned long long)g_presented,
s.numFramesActuallyPresented, (unsigned)s.status);
}
extern "C" NX2SL_API uint32_t nx2sl_frames_max(void) { return g_available == 1 ? g_max.load() : 0; }
extern "C" NX2SL_API uint64_t nx2sl_shown(void) { return g_shown; }
extern "C" NX2SL_API void nx2sl_shutdown(void) {
if (p_shutdown && g_available >= 0) p_shutdown();
}
Reference in New Issue View Git Blame Copy Permalink