diff --git a/docs/releases/v0.9.2.md b/docs/releases/v0.9.2.md new file mode 100644 index 00000000..db2c82ea --- /dev/null +++ b/docs/releases/v0.9.2.md @@ -0,0 +1,25 @@ +## Fixed + +- **A crash after switching FSR modes.** One FSR context served every mode, with another render size at each switch, + and the first frame after a switch from Quality to Native AA could end the game inside AMD's driver. Each render size + now gets a context of its own, as AMD's own sample makes them, and the one before goes a few frames later (a mode + switch takes about 50 ms more); with dynamic resolution on, one context made for changing sizes takes every step. +- **The light line at the bottom or right edge with FSR or DLSS, still seen in some frames.** The scene's outermost + rows and columns are now filled from the ones next to them in every frame the upscaler gets, not only after the + scene moved up or to the left. +- **A blocky picture for a moment whenever the camera swung quickly with FSR or DLSS.** A fast swing of the camera + around Link (11 to 18 degrees a frame) was taken for a cut to a new view, and the upscaler threw away what it had + built up at every frame of the swing; at Ultra Performance that showed the scene at a third of the picture's size. + Only a jump of the camera or a turn of 30 degrees or more in one frame counts as a cut now. + +## Known issues + +- At FSR's and DLSS's smaller modes, Link and other characters look coarser while they move: the upscaler gets only + the camera's motion for them, not their own, so it can keep less of what it built up there. +- Over water, a bright line can still show at the very bottom of the picture now and then with FSR or DLSS: the water + reads the scene before its last row is filled. +- Ultra Performance gains little frame rate over Quality with FSR 4. Most of a frame is the replay of the game's draws + on the CPU, which does not shrink with the render size, and FSR 4 itself takes 3.1 ms of the GPU's time at Ultra + Performance and 3.9 ms at Native AA on a Radeon RX 7900 XT (FSR 3.1: 0.5 ms). Measured standing in Hyrule Field at + 3416x1440: FSR 4 Native AA 52 fps, Quality 53, Ultra Performance 55; FSR 3.1 Native AA 55, Ultra Performance 66; + without an upscaler 70. diff --git a/runtime/vk_dlaa.c b/runtime/vk_dlaa.c index 3d58bd67..040db9be 100644 --- a/runtime/vk_dlaa.c +++ b/runtime/vk_dlaa.c @@ -1070,9 +1070,13 @@ static void cam_note(int view, int kept) { // that jumped: an area load or a spawn moves it 5 to 180 m in one frame. Such a frame now starts a new history, as // NVIDIA's DLSS guide asks for a camera cut: a move over 5 m, or a turn over 10 degrees together with a move over 0.5 m, // both per 60 Hz frame and scaled by the time since the last camera. Camera traces of play in the Fire Temple -// (NX2NAVI_MOUSE_TRACE) had a median of 0.05 m and 0.08 degrees a frame; the second test took only the area loads and a -// swing of the camera through a doorway at 12 to 18 degrees and 0.7 to 1.4 m a frame. A turn alone (the mouse camera, -// which orbits about 3 m from Link) does not count. Each cut is logged, up to 24. +// (NX2NAVI_MOUSE_TRACE) had a median of 0.05 m and 0.08 degrees a frame. A turn alone (the mouse camera, which orbits +// about 3 m from Link) does not count, and the turn with a move counts only from CUT_TURN degrees: a quick swing of the +// orbiting camera turns 11 to 18 degrees and moves 0.9 to 1.4 m a frame, and taken for a cut at 10 degrees it started a +// new history at every frame of the swing (nine in a row in Hyrule Field), which at FSR's or DLSS's smaller modes showed +// the scene at its render size for those frames, blocky and flickering. Its motion vectors carry it as well as any other +// turn. Each cut is logged, up to 24. +#define CUT_TURN 30.0 // v: the Ocarina scene context's rows 0-2, the view (fg_camera). static void cam_cut(const float *v) { static int on = -1; @@ -1091,7 +1095,7 @@ static void cam_cut(const float *v) { double k = (now - D.cut_ns) / 1e6 / (1000.0 / 60); // 60 Hz frames since the last camera if (k < 0.25) k = 0.25; if (k > 4) k = 4; - if (move > 5.0 * k || (turn > 10.0 * k && move > 0.5 * k)) { + if (move > 5.0 * k || (turn > CUT_TURN * k && move > 0.5 * k)) { D.reset = 1, D.cuts++; if (D.cut_logged < 24) D.cut_logged++, rt_log("dlaa: frame %llu: camera cut (moved %.2f m, turned %.1f degrees in %.1f ms): new " @@ -1816,35 +1820,45 @@ static void release_kept(const struct Size *making, int all) { making->w, making->h, all ? "after a release" : "larger inputs"); } -// The scene's last column and row after a jitter to the left or up. The jitter moves the scene's viewport by under half a -// pixel (vk_dlaa_viewport), which keeps every pixel centre inside it, yet on an RX 7900 XT (AMD's Windows driver) the -// draws of a viewport moved left left the last column out, and of one moved up the last row: the driver evidently limits -// drawing to the whole pixels inside the viewport. In a frame dump (NX2NAVI_DUMP, Hyrule Field, 3416x1440, FSR 4 Native -// AA) the depth's last row kept its clear value through every scene pass and the light target held the fog's colour -// there, which the upscaler blended into a grey line along the bottom and right edges of the picture that flickered with -// the jitter; with NX2NAVI_DLAA_JITTER=0, and without DLSS or FSR, there was no line. Before DLSS or FSR reads the scene, -// such a column or row of the light target and of the depth (which the motion vectors are made from) is filled from the -// one next to it. NX2NAVI_DLAA_EDGE_FILL=0 (test): not filled. -static void edge_fill(VkCommandBuffer cb, const VkExecTarget *t, const VkExecTarget *z, float jx, float jy) { +// The scene's outer columns and rows in a jittered frame. The jitter moves the scene's viewport by under half a pixel +// (vk_dlaa_viewport), which keeps every pixel centre inside it, yet on an RX 7900 XT (AMD's Windows driver) the draws of +// a viewport moved left left the last column out, and of one moved up the last row: the driver evidently limits drawing +// to the whole pixels inside the viewport. In a frame dump (NX2NAVI_DUMP, Hyrule Field, 3416x1440, FSR 4 Native AA) the +// depth's last row kept its clear value through every scene pass and the light target held the fog's colour there, +// which the upscaler blended into a light line along the bottom and right edges of the picture that flickered with the +// jitter; with NX2NAVI_DLAA_JITTER=0, and without DLSS or FSR, there was no line. Before DLSS or FSR reads the scene, the +// outermost column and row on each side of the light target and of the depth (which the motion vectors are made from) +// are filled from the ones next to them, in every jittered frame: filled only after a jitter to the left or up, the line +// stayed gone in every test run, yet a player's clip at Ultra Performance with HDR still showed the sky's colour in the +// last row of about half the frames, and another driver may round the near edge instead. A copied edge costs nothing +// visible: one render pixel at the picture's border, which the upscaler blends with its history. +// NX2NAVI_DLAA_EDGE_FILL=0 (test): not filled. +static void edge_fill(VkCommandBuffer cb, const VkExecTarget *t, const VkExecTarget *z, int jittered) { static int on = -1; if (on < 0) on = rt_knob("DLAA_EDGE_FILL", "1"); uint32_t w = t->width, h = t->height; - if (!on || w < 2 || h < 2 || (jx >= 0 && jy >= 0)) return; + if (!on || !jittered || w < 3 || h < 3) return; static int told; - if (!told++) rt_log("dlaa: the scene's last column or row after a jitter to the left or up is filled from the one before it"); - // the column first, then the row over the whole width, so the corner comes from the filled column + if (!told++) rt_log("dlaa: the scene's outer columns and rows are filled from the ones next to them in each jittered frame"); + // the columns first, then the rows over the whole width, so the corners come from the filled columns for (int pass = 0; pass < 2; pass++) { - if (pass ? jy >= 0 : jx >= 0) continue; - VkOffset3D from = pass ? (VkOffset3D){0, (int32_t)h - 2, 0} : (VkOffset3D){(int32_t)w - 2, 0, 0}; - VkOffset3D to = pass ? (VkOffset3D){0, (int32_t)h - 1, 0} : (VkOffset3D){(int32_t)w - 1, 0, 0}; - VkExtent3D ext = pass ? (VkExtent3D){w, 1, 1} : (VkExtent3D){1, h, 1}; + VkImageCopy r[2][2]; // [image][side] + for (int side = 0; side < 2; side++) { + int32_t end = pass ? (int32_t)h - 1 : (int32_t)w - 1, at = side ? end : 0, next = side ? end - 1 : 1; + VkOffset3D from = pass ? (VkOffset3D){0, next, 0} : (VkOffset3D){next, 0, 0}; + VkOffset3D to = pass ? (VkOffset3D){0, at, 0} : (VkOffset3D){at, 0, 0}; + VkExtent3D ext = pass ? (VkExtent3D){w, 1, 1} : (VkExtent3D){1, h, 1}; + for (int i = 0; i < 2; i++) { + const VkExecTarget *g = i ? z : t; + VkImageAspectFlags a = i ? VK_IMAGE_ASPECT_DEPTH_BIT : VK_IMAGE_ASPECT_COLOR_BIT; + r[i][side] = (VkImageCopy){.srcSubresource = {a, g->level, g->base_layer, 1}, .srcOffset = from, + .dstSubresource = {a, g->level, g->base_layer, 1}, .dstOffset = to, .extent = ext}; + } + } for (int i = 0; i < 2; i++) { const VkExecTarget *g = i ? z : t; - if (!g->image || g->width != w || g->height != h) continue; - VkImageAspectFlags a = i ? VK_IMAGE_ASPECT_DEPTH_BIT : VK_IMAGE_ASPECT_COLOR_BIT; - VkImageCopy r = {.srcSubresource = {a, g->level, g->base_layer, 1}, .srcOffset = from, - .dstSubresource = {a, g->level, g->base_layer, 1}, .dstOffset = to, .extent = ext}; - vkCmdCopyImage(cb, g->image, VK_IMAGE_LAYOUT_GENERAL, g->image, VK_IMAGE_LAYOUT_GENERAL, 1, &r); + if (g->image && g->width == w && g->height == h) + vkCmdCopyImage(cb, g->image, VK_IMAGE_LAYOUT_GENERAL, g->image, VK_IMAGE_LAYOUT_GENERAL, 2, r[i]); } full_barrier(cb); } @@ -1880,7 +1894,7 @@ static void evaluate(VkExec *ex, const VkExecTarget *dst) { VkImageView zv = vk_exec_target_view(z->img, z->format, VK_IMAGE_VIEW_TYPE_2D, VK_IMAGE_ASPECT_DEPTH_BIT, z->level, z->base_layer, 1, 0); float jx = D.jit_used ? D.ux : 0.0f, jy = D.jit_used ? D.uy : 0.0f; - edge_fill(cb, t, z, jx, jy); // before the motion vectors read the depth + edge_fill(cb, t, z, D.jit_used); // before the motion vectors read the depth static int no_mv = -1; if (no_mv < 0) no_mv = !rt_knob("DLAA_MV", "1"); // NX2NAVI_DLAA_MV=0: zero motion (test) int masked = 0, mv = no_mv ? 0 : mv_pass(ex, e, zv, jx, jy, char_bias(), &masked);