PolytreeDE/picom
1
0
Fork 0
mirror of https://github.com/yshui/picom.git synced 2024-12-09 14:18:57 -05:00
Code Releases Activity

Merge pull request #968 from yshui/pacing

Browse source
This commit is contained in:
Yuxuan Shui 2023-06-12 11:46:17 +01:00 committed by GitHub
commit 608423a5ce
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
21 changed files with 817 additions and 154 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

3
man/picom.1.asciidoc
View file

@ -109,6 +109,9 @@ OPTIONS
*--rounded-corners-exclude* 'CONDITION'::
Exclude conditions for rounded corners.
*--no-frame-pacing*::
Disable vsync-aware frame pacing. By default, the compositor tries to make sure it only renders once per vblank interval, and also the render happens as late as possible to minimize the latency from updates to the screen. However this can sometimes cause stuttering, or even lowered frame rate. This option can be used to disable frame pacing.
*--mark-wmwin-focused*::
Try to detect WM windows (a non-override-redirect window with no child that has 'WM_STATE') and mark them as active.

24
src/backend/backend.c
View file

@ -81,7 +81,7 @@ void handle_device_reset(session_t *ps) {
}
/// paint all windows
void paint_all_new(session_t *ps, struct managed_win *t, bool ignore_damage) {
void paint_all_new(session_t *ps, struct managed_win *t) {
if (ps->backend_data->ops->device_status &&
ps->backend_data->ops->device_status(ps->backend_data) != DEVICE_STATUS_NORMAL) {
return handle_device_reset(ps);
@ -100,12 +100,7 @@ void paint_all_new(session_t *ps, struct managed_win *t, bool ignore_damage) {
// the paints bleed out of the damage region, it will destroy
// part of the image we want to reuse
region_t reg_damage;
if (!ignore_damage) {
reg_damage = get_damage(ps, ps->o.monitor_repaint || !ps->o.use_damage);
} else {
pixman_region32_init(&reg_damage);
pixman_region32_copy(&reg_damage, &ps->screen_reg);
}
reg_damage = get_damage(ps, ps->o.monitor_repaint || !ps->o.use_damage);
if (!pixman_region32_not_empty(&reg_damage)) {
pixman_region32_fini(&reg_damage);
@ -181,6 +176,21 @@ void paint_all_new(session_t *ps, struct managed_win *t, bool ignore_damage) {
region_t reg_shadow_clip;
pixman_region32_init(&reg_shadow_clip);
struct timespec now;
clock_gettime(CLOCK_MONOTONIC, &now);
auto now_us = (uint64_t)(now.tv_sec * 1000000 + now.tv_nsec / 1000);
if (ps->next_render > 0) {
log_trace("Render schedule deviation: %ld us (%s) %ld %ld",
labs((int64_t)now_us - (int64_t)ps->next_render),
now_us < ps->next_render ? "early" : "late", now_us,
ps->next_render);
ps->last_schedule_delay = 0;
if (now_us > ps->next_render) {
ps->last_schedule_delay = now_us - ps->next_render;
}
}
ps->did_render = true;
if (ps->backend_data->ops->prepare) {
ps->backend_data->ops->prepare(ps->backend_data, &reg_paint);
}

11
src/backend/backend.h
View file

@ -292,6 +292,14 @@ struct backend_operations {
/// Optional
int (*buffer_age)(backend_t *backend_data);
/// Get the render time of the last frame. If the render is still in progress,
/// returns false. The time is returned in `ts`. Frames are delimited by the
/// present() calls. i.e. after a present() call, last_render_time() should start
/// reporting the time of the just presen1ted frame.
///
/// Optional, if not available, the most conservative estimation will be used.
bool (*last_render_time)(backend_t *backend_data, struct timespec *ts);
/// The maximum number buffer_age might return.
int max_buffer_age;
@ -363,5 +371,4 @@ struct backend_operations {
extern struct backend_operations *backend_list[];
void paint_all_new(session_t *ps, struct managed_win *const t, bool ignore_damage)
attr_nonnull(1);
void paint_all_new(session_t *ps, struct managed_win *const t) attr_nonnull(1);

2
src/backend/driver.c
View file

@ -15,8 +15,6 @@
/// Apply driver specified global workarounds. It's safe to call this multiple times.
void apply_driver_workarounds(struct session *ps, enum driver driver) {
if (driver & DRIVER_NVIDIA) {
// setenv("__GL_YIELD", "usleep", true);
setenv("__GL_MaxFramesAllowed", "1", true);
ps->o.xrender_sync_fence = true;
}
}

5
src/backend/gl/egl.c
View file

@ -320,9 +320,6 @@ static void egl_present(backend_t *base, const region_t *region attr_unused) {
struct egl_data *gd = (void *)base;
gl_present(base, region);
eglSwapBuffers(gd->display, gd->target_win);
if (!gd->gl.is_nvidia) {
glFinish();
}
}
static int egl_buffer_age(backend_t *base) {
@ -372,6 +369,7 @@ struct backend_operations egl_ops = {
.deinit = egl_deinit,
.bind_pixmap = egl_bind_pixmap,
.release_image = gl_release_image,
.prepare = gl_prepare,
.compose = gl_compose,
.image_op = gl_image_op,
.set_image_property = gl_set_image_property,
@ -380,6 +378,7 @@ struct backend_operations egl_ops = {
.is_image_transparent = default_is_image_transparent,
.present = egl_present,
.buffer_age = egl_buffer_age,
.last_render_time = gl_last_render_time,
.create_shadow_context = gl_create_shadow_context,
.destroy_shadow_context = gl_destroy_shadow_context,
.render_shadow = backend_render_shadow_from_mask,

33
src/backend/gl/gl_common.c
View file

@ -22,6 +22,11 @@
#include "backend/backend_common.h"
#include "backend/gl/gl_common.h"
void gl_prepare(backend_t *base, const region_t *reg attr_unused) {
auto gd = (struct gl_data *)base;
glBeginQuery(GL_TIME_ELAPSED, gd->frame_timing[gd->current_frame_timing]);
}
GLuint gl_create_shader(GLenum shader_type, const char *shader_str) {
log_trace("===\n%s\n===", shader_str);
@ -800,6 +805,9 @@ uint64_t gl_get_shader_attributes(backend_t *backend_data attr_unused, void *sha
}
bool gl_init(struct gl_data *gd, session_t *ps) {
glGenQueries(2, gd->frame_timing);
gd->current_frame_timing = 0;
// Initialize GLX data structure
glDisable(GL_DEPTH_TEST);
glDepthMask(GL_FALSE);
@ -945,7 +953,7 @@ bool gl_init(struct gl_data *gd, session_t *ps) {
const char *vendor = (const char *)glGetString(GL_VENDOR);
log_debug("GL_VENDOR = %s", vendor);
if (strcmp(vendor, "NVIDIA Corporation") == 0) {
log_info("GL vendor is NVIDIA, don't use glFinish");
log_info("GL vendor is NVIDIA, enable xrender sync fence.");
gd->is_nvidia = true;
} else {
gd->is_nvidia = false;
@ -1154,10 +1162,33 @@ void gl_present(backend_t *base, const region_t *region) {
glDeleteBuffers(2, bo);
glDeleteVertexArrays(1, &vao);
glEndQuery(GL_TIME_ELAPSED);
gd->current_frame_timing ^= 1;
gl_check_err();
free(coord);
free(indices);
}
bool gl_last_render_time(backend_t *base, struct timespec *ts) {
auto gd = (struct gl_data *)base;
GLint available = 0;
glGetQueryObjectiv(gd->frame_timing[gd->current_frame_timing ^ 1],
GL_QUERY_RESULT_AVAILABLE, &available);
if (!available) {
return false;
}
GLuint64 time;
glGetQueryObjectui64v(gd->frame_timing[gd->current_frame_timing ^ 1],
GL_QUERY_RESULT, &time);
ts->tv_sec = (long)(time / 1000000000);
ts->tv_nsec = (long)(time % 1000000000);
gl_check_err();
return true;
}
bool gl_image_op(backend_t *base, enum image_operations op, void *image_data,
const region_t *reg_op, const region_t *reg_visible attr_unused, void *arg) {
struct backend_image *tex = image_data;

4
src/backend/gl/gl_common.h
View file

@ -108,6 +108,8 @@ struct gl_data {
gl_shadow_shader_t shadow_shader;
GLuint back_texture, back_fbo;
GLint back_format;
GLuint frame_timing[2];
int current_frame_timing;
GLuint present_prog;
bool dithered_present;
@ -129,6 +131,7 @@ typedef struct session session_t;
#define GL_PROG_MAIN_INIT \
{ .prog = 0, .unifm_opacity = -1, .unifm_invert_color = -1, .unifm_tex = -1, }
void gl_prepare(backend_t *base, const region_t *reg);
void x_rect_to_coords(int nrects, const rect_t *rects, coord_t image_dst,
int extent_height, int texture_height, int root_height,
bool y_inverted, GLint *coord, GLuint *indices);
@ -142,6 +145,7 @@ void gl_destroy_window_shader(backend_t *backend_data, void *shader);
uint64_t gl_get_shader_attributes(backend_t *backend_data, void *shader);
bool gl_set_image_property(backend_t *backend_data, enum image_properties prop,
void *image_data, void *args);
bool gl_last_render_time(backend_t *backend_data, struct timespec *time);
/**
* @brief Render a region with texture data.

5
src/backend/gl/glx.c
View file

@ -468,9 +468,6 @@ static void glx_present(backend_t *base, const region_t *region attr_unused) {
struct _glx_data *gd = (void *)base;
gl_present(base, region);
glXSwapBuffers(gd->display, gd->target_win);
if (!gd->gl.is_nvidia) {
glFinish();
}
}
static int glx_buffer_age(backend_t *base) {
@ -528,6 +525,7 @@ struct backend_operations glx_ops = {
.deinit = glx_deinit,
.bind_pixmap = glx_bind_pixmap,
.release_image = gl_release_image,
.prepare = gl_prepare,
.compose = gl_compose,
.image_op = gl_image_op,
.set_image_property = gl_set_image_property,
@ -536,6 +534,7 @@ struct backend_operations glx_ops = {
.is_image_transparent = default_is_image_transparent,
.present = glx_present,
.buffer_age = glx_buffer_age,
.last_render_time = gl_last_render_time,
.create_shadow_context = gl_create_shadow_context,
.destroy_shadow_context = gl_destroy_shadow_context,
.render_shadow = backend_render_shadow_from_mask,

29
src/common.h
View file

@ -58,6 +58,7 @@
#include "list.h"
#include "region.h"
#include "render.h"
#include "statistics.h"
#include "types.h"
#include "utils.h"
#include "win_defs.h"
@ -156,9 +157,8 @@ typedef struct session {
ev_timer unredir_timer;
/// Timer for fading
ev_timer fade_timer;
/// Use an ev_idle callback for drawing
/// So we only start drawing when events are processed
ev_idle draw_idle;
/// Use an ev_timer callback for drawing
ev_timer draw_timer;
/// Called every time we have timeouts or new data on socket,
/// so we can be sure if xcb read from X socket at anytime during event
/// handling, we will not left any event unhandled in the queue
@ -240,6 +240,29 @@ typedef struct session {
bool first_frame;
/// Whether screen has been turned off
bool screen_is_off;
/// Event context for X Present extension.
uint32_t present_event_id;
xcb_special_event_t *present_event;
/// When last MSC event happened, in useconds.
uint64_t last_msc_instant;
/// The last MSC number
uint64_t last_msc;
/// When the currently rendered frame will be displayed.
/// 0 means there is no pending frame.
uint64_t target_msc;
/// The delay between when the last frame was scheduled to be rendered, and when
/// the render actually started.
uint64_t last_schedule_delay;
/// When do we want our next frame to start rendering.
uint64_t next_render;
/// Did we actually render the last frame. Sometimes redraw will be scheduled only
/// to find out nothing has changed. In which case this will be set to false.
bool did_render;
/// Whether we can perform frame pacing.
bool frame_pacing;
/// Render statistics
struct render_statistics render_stats;
// === Operation related ===
/// Flags related to the root window

1
src/config.c
View file

@ -742,6 +742,7 @@ char *parse_config(options_t *opt, const char *config_file, bool *shadow_enable,
.logpath = NULL,
.use_damage = true,
.no_frame_pacing = false,
.shadow_red = 0.0,
.shadow_green = 0.0,

2
src/config.h
View file

@ -140,6 +140,8 @@ typedef struct options {
bool vsync_use_glfinish;
/// Whether use damage information to help limit the area to paint
bool use_damage;
/// Disable frame pacing
bool no_frame_pacing;
// === Shadow ===
/// Red, green and blue tone of the shadow.

3
src/config_libconfig.c
View file

@ -363,6 +363,9 @@ char *parse_config_libconfig(options_t *opt, const char *config_file, bool *shad
// --corner-radius-rules
parse_cfg_condlst_corner(opt, &cfg, "corner-radius-rules");
// --no-frame-pacing
lcfg_lookup_bool(&cfg, "no-frame-pacing", &opt->no_frame_pacing);
// -e (frame_opacity)
config_lookup_float(&cfg, "frame-opacity", &opt->frame_opacity);
// -c (shadow_enable)

2
src/meson.build
View file

@ -9,7 +9,7 @@ base_deps = [
srcs = [ files('picom.c', 'win.c', 'c2.c', 'x.c', 'config.c', 'vsync.c', 'utils.c',
'diagnostic.c', 'string_utils.c', 'render.c', 'kernel.c', 'log.c',
'options.c', 'event.c', 'cache.c', 'atom.c', 'file_watch.c') ]
'options.c', 'event.c', 'cache.c', 'atom.c', 'file_watch.c', 'statistics.c') ]
picom_inc = include_directories('.')
cflags = []

2
src/options.c
View file

@ -176,6 +176,7 @@ static const struct picom_option picom_options[] = {
"rendered screen. Reduces banding artifacts, but might cause performance "
"degradation. Only works with OpenGL."},
// 340 is corner-radius-rules
{"no-frame-pacing" , no_argument , 341, NULL , "Disable frame pacing. This might increase the latency."},
{"legacy-backends" , no_argument , 733, NULL , "Use deprecated version of the backends."},
{"monitor-repaint" , no_argument , 800, NULL , "Highlight the updated area of the screen. For debugging."},
{"diagnostics" , no_argument , 801, NULL , "Print diagnostic information"},
@ -738,6 +739,7 @@ bool get_cfg(options_t *opt, int argc, char *const *argv, bool shadow_enable,
// --dithered-present
opt->dithered_present = true;
break;
P_CASEBOOL(341, no_frame_pacing);
P_CASEBOOL(733, legacy_backends);
P_CASEBOOL(800, monitor_repaint);
case 801:

377
src/picom.c
View file

@ -16,8 +16,10 @@
#include <errno.h>
#include <fcntl.h>
#include <inttypes.h>
#include <stddef.h>
#include <stdio.h>
#include <string.h>
#include <sys/resource.h>
#include <unistd.h>
#include <xcb/composite.h>
#include <xcb/damage.h>
@ -43,7 +45,6 @@
#endif
#include "backend/backend.h"
#include "c2.h"
#include "config.h"
#include "diagnostic.h"
#include "log.h"
#include "region.h"
@ -60,6 +61,7 @@
#include "file_watch.h"
#include "list.h"
#include "options.h"
#include "statistics.h"
#include "uthash_extra.h"
/// Get session_t pointer from a pointer to a member of session_t
@ -133,6 +135,7 @@ void check_dpms_status(EV_P attr_unused, ev_timer *w, int revents attr_unused) {
}
auto now_screen_is_off = dpms_screen_is_off(r);
if (ps->screen_is_off != now_screen_is_off) {
log_debug("Screen is now %s", now_screen_is_off ? "off" : "on");
ps->screen_is_off = now_screen_is_off;
queue_redraw(ps);
}
@ -145,21 +148,173 @@ void check_dpms_status(EV_P attr_unused, ev_timer *w, int revents attr_unused) {
* XXX move to win.c
*/
static inline struct managed_win *find_win_all(session_t *ps, const xcb_window_t wid) {
if (!wid || PointerRoot == wid || wid == ps->root || wid == ps->overlay)
if (!wid || PointerRoot == wid || wid == ps->root || wid == ps->overlay) {
return NULL;
}
auto w = find_managed_win(ps, wid);
if (!w)
if (!w) {
w = find_toplevel(ps, wid);
if (!w)
}
if (!w) {
w = find_managed_window_or_parent(ps, wid);
}
return w;
}
/// How many seconds into the future should we start rendering the next frame.
///
/// Renders are scheduled like this:
///
/// 1. queue_redraw() registers the intention to render. redraw_needed is set to true to
/// indicate what is on screen needs to be updated.
/// 2. then, we need to figure out the best time to start rendering. first, we need to
/// know when the next frame will be displayed on screen. we have this information from
/// the Present extension: we know when was the last frame displayed, and we know the
/// refresh rate. so we can calculate the next frame's display time. if our render time
/// estimation shows we could miss that target, we push the target back one frame.
/// 3. if there is already render completed for that target frame, or there is a render
/// currently underway, we don't do anything, and wait for the next Present Complete
/// Notify event to try to schedule again.
/// 4. otherwise, we schedule a render for that target frame. we use past statistics about
/// how long our renders took to figure out when to start rendering. we start rendering
/// at the latest point of time possible to still hit the target frame.
///
/// The `triggered_by_timer` parameter is used to indicate whether this function is
/// triggered by a steady timer, i.e. we are rendering for each vblank. The other case is
/// when we stop rendering for a while because there is no changes on screen, then
/// something changed and schedule_render is triggered by a DamageNotify. The idea is that
/// when the schedule is triggered by a steady timer, schedule_render will be called at a
/// predictable offset into each vblank.
void schedule_render(session_t *ps, bool triggered_by_vblank) {
double delay_s = 0;
ps->next_render = 0;
if (!ps->frame_pacing || !ps->redirected) {
// Not doing frame pacing, schedule a render immediately, if not already
// scheduled.
// If not redirected, we schedule immediately to have a chance to
// redirect. We won't have frame or render timing information anyway.
if (!ev_is_active(&ps->draw_timer)) {
// We don't know the msc, so we set it to 1, because 0 is a
// special value
ps->target_msc = 1;
goto schedule;
}
return;
}
struct timespec render_time;
bool completed =
ps->backend_data->ops->last_render_time(ps->backend_data, &render_time);
if (!completed || ev_is_active(&ps->draw_timer)) {
// There is already a render underway (either just scheduled, or is
// rendered but awaiting completion), don't schedule another one.
if (ps->target_msc <= ps->last_msc) {
log_debug("Target frame %ld is in the past, but we are still "
"rendering",
ps->target_msc);
// We missed our target, push it back one frame
ps->target_msc = ps->last_msc + 1;
}
log_trace("Still rendering for target frame %ld, not scheduling another "
"render",
ps->target_msc);
return;
}
if (ps->target_msc > ps->last_msc) {
// Render for the target frame is completed, but is yet to be displayed.
// Don't schedule another render.
log_trace("Target frame %ld is in the future, and we have already "
"rendered, last msc: %d",
ps->target_msc, (int)ps->last_msc);
return;
}
struct timespec now;
clock_gettime(CLOCK_MONOTONIC, &now);
auto now_us = (uint64_t)now.tv_sec * 1000000 + (uint64_t)now.tv_nsec / 1000;
if (triggered_by_vblank) {
log_trace("vblank schedule delay: %ld us", now_us - ps->last_msc_instant);
}
int render_time_us =
(int)(render_time.tv_sec * 1000000L + render_time.tv_nsec / 1000L);
if (ps->target_msc == ps->last_msc) {
// The frame has just been displayed, record its render time;
if (ps->did_render) {
log_trace("Last render call took: %d (gpu) + %d (cpu) us, "
"last_msc: %" PRIu64,
render_time_us, (int)ps->last_schedule_delay, ps->last_msc);
render_statistics_add_render_time_sample(
&ps->render_stats, render_time_us + (int)ps->last_schedule_delay);
}
ps->target_msc = 0;
ps->did_render = false;
ps->last_schedule_delay = 0;
}
unsigned int divisor = 0;
auto render_budget = render_statistics_get_budget(&ps->render_stats, &divisor);
auto frame_time = render_statistics_get_vblank_time(&ps->render_stats);
if (frame_time == 0) {
// We don't have enough data for render time estimates, maybe there's
// no frame rendered yet, or the backend doesn't support render timing
// information, schedule render immediately.
ps->target_msc = ps->last_msc + 1;
goto schedule;
}
const auto deadline = ps->last_msc_instant + (unsigned long)divisor * frame_time;
unsigned int available = 0;
if (deadline > now_us) {
available = (unsigned int)(deadline - now_us);
}
ps->target_msc = ps->last_msc + divisor;
if (available > render_budget) {
delay_s = (double)(available - render_budget) / 1000000.0;
ps->next_render = deadline - render_budget;
} else {
delay_s = 0;
ps->next_render = now_us;
}
if (delay_s > 1) {
log_warn("Delay too long: %f s, render_budget: %d us, frame_time: "
"%" PRIu32 " us, now_us: %" PRIu64 " us, next_msc: %" PRIu64 " u"
"s",
delay_s, render_budget, frame_time, now_us, deadline);
}
log_trace("Delay: %.6lf s, last_msc: %" PRIu64 ", render_budget: %d, frame_time: "
"%" PRIu32 ", now_us: %" PRIu64 ", next_msc: %" PRIu64 ", "
"target_msc: %" PRIu64 ", divisor: %d",
delay_s, ps->last_msc_instant, render_budget, frame_time, now_us,
deadline, ps->target_msc, divisor);
schedule:
assert(!ev_is_active(&ps->draw_timer));
ev_timer_set(&ps->draw_timer, delay_s, 0);
ev_timer_start(ps->loop, &ps->draw_timer);
}
void queue_redraw(session_t *ps) {
if (ps->screen_is_off) {
// The screen is off, if there is a draw queued for the next frame (i.e.
// ps->redraw_needed == true), it won't be triggered until the screen is
// on again, because the abnormal Present events we will receive from the
// X server when the screen is off. Yet we need the draw_callback to be
// called as soon as possible so the screen can be unredirected.
// So here we unconditionally start the draw timer.
ev_timer_stop(ps->loop, &ps->draw_timer);
ev_timer_set(&ps->draw_timer, 0, 0);
ev_timer_start(ps->loop, &ps->draw_timer);
return;
}
// Whether we have already rendered for the current frame.
// If frame pacing is not enabled, pretend this is false.
// If --benchmark is used, redraw is always queued
if (!ps->redraw_needed && !ps->o.benchmark) {
ev_idle_start(ps->loop, &ps->draw_idle);
schedule_render(ps, false);
}
ps->redraw_needed = true;
}
@ -635,7 +790,6 @@ static void configure_root(session_t *ps) {
}
force_repaint(ps);
}
return;
}
static void handle_root_flags(session_t *ps) {
@ -1291,6 +1445,39 @@ static bool redirect_start(session_t *ps) {
pixman_region32_init(&ps->damage_ring[i]);
}
ps->frame_pacing = !ps->o.no_frame_pacing;
if ((ps->o.legacy_backends || ps->o.benchmark || !ps->backend_data->ops->last_render_time) &&
ps->frame_pacing) {
// Disable frame pacing if we are using a legacy backend or if we are in
// benchmark mode, or if the backend doesn't report render time
log_info("Disabling frame pacing.");
ps->frame_pacing = false;
}
if (ps->present_exists && ps->frame_pacing) {
ps->present_event_id = x_new_id(ps->c);
auto select_input = xcb_present_select_input(
ps->c, ps->present_event_id, session_get_target_window(ps),
XCB_PRESENT_EVENT_MASK_COMPLETE_NOTIFY);
auto notify_msc =
xcb_present_notify_msc(ps->c, session_get_target_window(ps), 0, 0, 1, 0);
set_cant_fail_cookie(ps, select_input);
set_cant_fail_cookie(ps, notify_msc);
ps->present_event = xcb_register_for_special_xge(
ps->c, &xcb_present_id, ps->present_event_id, NULL);
// Initialize rendering and frame timing statistics, and frame pacing
// states.
ps->last_msc_instant = 0;
ps->last_msc = 0;
ps->last_schedule_delay = 0;
ps->target_msc = 0;
render_statistics_reset(&ps->render_stats);
} else if (ps->frame_pacing) {
log_error("Present extension is not supported, frame pacing disabled.");
ps->frame_pacing = false;
}
// Must call XSync() here
x_sync(ps->c);
@ -1332,6 +1519,14 @@ static void unredirect(session_t *ps) {
free(ps->damage_ring);
ps->damage_ring = ps->damage = NULL;
if (ps->present_event_id) {
xcb_present_select_input(ps->c, ps->present_event_id,
session_get_target_window(ps), 0);
ps->present_event_id = XCB_NONE;
xcb_unregister_for_special_event(ps->c, ps->present_event);
ps->present_event = NULL;
}
// Must call XSync() here
x_sync(ps->c);
@ -1339,9 +1534,93 @@ static void unredirect(session_t *ps) {
log_debug("Screen unredirected.");
}
static void
handle_present_complete_notify(session_t *ps, xcb_present_complete_notify_event_t *cne) {
if (cne->kind != XCB_PRESENT_COMPLETE_KIND_NOTIFY_MSC) {
return;
}
bool event_is_invalid = false;
if (ps->frame_pacing) {
auto next_msc = cne->msc + 1;
if (cne->msc <= ps->last_msc || cne->ust == 0) {
// X sometimes sends duplicate/bogus MSC events, don't
// use the msc value. Also ignore these events.
//
// See:
// https://gitlab.freedesktop.org/xorg/xserver/-/issues/1418
next_msc = ps->last_msc + 1;
event_is_invalid = true;
}
auto cookie = xcb_present_notify_msc(ps->c, session_get_target_window(ps),
0, next_msc, 0, 0);
set_cant_fail_cookie(ps, cookie);
}
if (event_is_invalid) {
return;
}
struct timespec now;
clock_gettime(CLOCK_MONOTONIC, &now);
uint64_t now_usec = (uint64_t)(now.tv_sec * 1000000 + now.tv_nsec / 1000);
uint64_t drift;
if (cne->ust > now_usec) {
drift = cne->ust - now_usec;
} else {
drift = now_usec - cne->ust;
}
if (ps->last_msc_instant != 0) {
auto frame_count = cne->msc - ps->last_msc;
int frame_time = (int)((cne->ust - ps->last_msc_instant) / frame_count);
render_statistics_add_vblank_time_sample(&ps->render_stats, frame_time);
log_trace("Frame count %lu, frame time: %d us, rolling average: %u us, "
"msc: %" PRIu64 ", offset: %d us",
frame_count, frame_time,
render_statistics_get_vblank_time(&ps->render_stats), cne->ust,
(int)drift);
} else if (drift > 1000000 && ps->frame_pacing) {
// This is the first MSC event we receive, let's check if the timestamps
// align with the monotonic clock. If not, disable frame pacing because we
// can't schedule frames reliably.
log_error("Temporal anomaly detected, frame pacing disabled. (Are we "
"running inside a time namespace?), %" PRIu64 " %" PRIu64,
now_usec, ps->last_msc_instant);
ps->frame_pacing = false;
}
ps->last_msc_instant = cne->ust;
ps->last_msc = cne->msc;
if (ps->redraw_needed) {
schedule_render(ps, true);
}
}
static void handle_present_events(session_t *ps) {
if (!ps->present_event) {
// Screen not redirected
return;
}
xcb_present_generic_event_t *ev;
while ((ev = (void *)xcb_poll_for_special_event(ps->c, ps->present_event))) {
if (ev->event != ps->present_event_id) {
// This event doesn't have the right event context, it's not meant
// for us.
goto next;
}
// We only subscribed to the complete notify event.
assert(ev->evtype == XCB_PRESENT_EVENT_COMPLETE_NOTIFY);
handle_present_complete_notify(ps, (void *)ev);
next:
free(ev);
}
}
// Handle queued events before we go to sleep
static void handle_queued_x_events(EV_P attr_unused, ev_prepare *w, int revents attr_unused) {
session_t *ps = session_ptr(w, event_check);
handle_present_events(ps);
xcb_generic_event_t *ev;
while ((ev = xcb_poll_for_queued_event(ps->c))) {
ev_handle(ps, ev);
@ -1403,6 +1682,8 @@ static void tmout_unredir_callback(EV_P attr_unused, ev_timer *w, int revents at
}
static void fade_timer_callback(EV_P attr_unused, ev_timer *w, int revents attr_unused) {
// TODO(yshui): do we still need the fade timer? we queue redraw automatically in
// draw_callback_impl if animation is running.
session_t *ps = session_ptr(w, fade_timer);
queue_redraw(ps);
}
@ -1460,8 +1741,24 @@ static void handle_pending_updates(EV_P_ struct session *ps) {
}
static void draw_callback_impl(EV_P_ session_t *ps, int revents attr_unused) {
struct timespec now;
int64_t draw_callback_enter_us;
clock_gettime(CLOCK_MONOTONIC, &now);
draw_callback_enter_us = (now.tv_sec * 1000000LL + now.tv_nsec / 1000);
if (ps->next_render != 0) {
log_trace("Schedule delay: %" PRIi64 " us",
draw_callback_enter_us - (int64_t)ps->next_render);
}
handle_pending_updates(EV_A_ ps);
int64_t after_handle_pending_updates_us;
clock_gettime(CLOCK_MONOTONIC, &now);
after_handle_pending_updates_us = (now.tv_sec * 1000000LL + now.tv_nsec / 1000);
log_trace("handle_pending_updates took: %" PRIi64 " us",
after_handle_pending_updates_us - draw_callback_enter_us);
if (ps->first_frame) {
// If we are still rendering the first frame, if some of the windows are
// unmapped/destroyed during the above handle_pending_updates() call, they
@ -1518,15 +1815,21 @@ static void draw_callback_impl(EV_P_ session_t *ps, int revents attr_unused) {
ev_timer_start(EV_A_ & ps->fade_timer);
}
int64_t after_preprocess_us;
clock_gettime(CLOCK_MONOTONIC, &now);
after_preprocess_us = (now.tv_sec * 1000000LL + now.tv_nsec / 1000);
log_trace("paint_preprocess took: %" PRIi64 " us",
after_preprocess_us - after_handle_pending_updates_us);
// If the screen is unredirected, free all_damage to stop painting
if (ps->redirected && ps->o.stoppaint_force != ON) {
static int paint = 0;
log_trace("Render start, frame %d", paint);
if (!ps->o.legacy_backends) {
paint_all_new(ps, bottom, false);
paint_all_new(ps, bottom);
} else {
paint_all(ps, bottom, false);
paint_all(ps, bottom);
}
log_trace("Render end");
@ -1544,18 +1847,21 @@ static void draw_callback_impl(EV_P_ session_t *ps, int revents attr_unused) {
// TODO(yshui) Investigate how big the X critical section needs to be. There are
// suggestions that rendering should be in the critical section as well.
// Queue redraw if animation is running. This should be picked up by next present
// event.
ps->redraw_needed = animation;
}
static void draw_callback(EV_P_ ev_idle *w, int revents) {
session_t *ps = session_ptr(w, draw_idle);
static void draw_callback(EV_P_ ev_timer *w, int revents) {
session_t *ps = session_ptr(w, draw_timer);
draw_callback_impl(EV_A_ ps, revents);
ev_timer_stop(EV_A_ w);
// Don't do painting non-stop unless we are in benchmark mode, or if
// draw_callback_impl thinks we should continue painting.
if (!ps->o.benchmark && !ps->redraw_needed) {
ev_idle_stop(EV_A_ & ps->draw_idle);
// Immediately start next frame if we are in benchmark mode.
if (ps->o.benchmark) {
ev_timer_set(w, 0, 0);
ev_timer_start(EV_A_ w);
}
}
@ -1697,6 +2003,8 @@ static session_t *session_init(int argc, char **argv, Display *dpy,
.white_picture = XCB_NONE,
.shadow_context = NULL,
.last_msc = 0,
#ifdef CONFIG_VSYNC_DRM
.drm_fd = -1,
#endif
@ -1716,6 +2024,7 @@ static session_t *session_init(int argc, char **argv, Display *dpy,
.randr_exists = 0,
.randr_event = 0,
.randr_error = 0,
.present_event_id = XCB_NONE,
.glx_exists = false,
.glx_event = 0,
.glx_error = 0,
@ -1743,6 +2052,9 @@ static session_t *session_init(int argc, char **argv, Display *dpy,
ps->loop = EV_DEFAULT;
pixman_region32_init(&ps->screen_reg);
// TODO(yshui) investigate what's the best window size
render_statistics_init(&ps->render_stats, 128);
ps->pending_reply_tail = &ps->pending_reply_head;
ps->o.show_all_xerrors = all_xerrors;
@ -2162,7 +2474,7 @@ static session_t *session_init(int argc, char **argv, Display *dpy,
ev_io_init(&ps->xiow, x_event_callback, ConnectionNumber(ps->dpy), EV_READ);
ev_io_start(ps->loop, &ps->xiow);
ev_init(&ps->unredir_timer, tmout_unredir_callback);
ev_idle_init(&ps->draw_idle, draw_callback);
ev_init(&ps->draw_timer, draw_callback);
ev_init(&ps->fade_timer, fade_timer_callback);
@ -2264,6 +2576,31 @@ err:
free(ps);
return NULL;
}
void set_rr_scheduling(void) {
struct rlimit rlim;
if (getrlimit(RLIMIT_RTPRIO, &rlim) != 0) {
log_warn("Failed to get RLIMIT_RTPRIO, not setting real-time priority");
return;
}
int old_policy;
int ret;
struct sched_param param;
ret = pthread_getschedparam(pthread_self(), &old_policy, &param);
if (ret != 0) {
log_debug("Failed to get old scheduling priority");
return;
}
param.sched_priority = (int)rlim.rlim_cur;
ret = pthread_setschedparam(pthread_self(), SCHED_RR, &param);
if (ret != 0) {
log_info("Failed to set scheduling priority to %lu", rlim.rlim_cur);
return;
}
log_info("Set scheduling priority to %lu", rlim.rlim_cur);
}
/**
* Destroy a session.
@ -2378,6 +2715,8 @@ static void session_destroy(session_t *ps) {
free(ps->o.glx_fshader_win_str);
x_free_randr_info(ps);
render_statistics_destroy(&ps->render_stats);
// Release custom window shaders
free(ps->o.window_shader_fg);
struct shader_info *shader, *tmp;
@ -2457,7 +2796,7 @@ static void session_destroy(session_t *ps) {
ev_timer_stop(ps->loop, &ps->unredir_timer);
ev_timer_stop(ps->loop, &ps->fade_timer);
ev_timer_stop(ps->loop, &ps->dpms_check_timer);
ev_idle_stop(ps->loop, &ps->draw_idle);
ev_timer_stop(ps->loop, &ps->draw_timer);
ev_prepare_stop(ps->loop, &ps->event_check);
ev_signal_stop(ps->loop, &ps->usr1_signal);
ev_signal_stop(ps->loop, &ps->int_signal);
@ -2469,9 +2808,11 @@ static void session_destroy(session_t *ps) {
* @param ps current session
*/
static void session_run(session_t *ps) {
// In benchmark mode, we want draw_idle handler to always be active
set_rr_scheduling();
// In benchmark mode, we want draw_timer handler to always be active
if (ps->o.benchmark) {
ev_idle_start(ps->loop, &ps->draw_idle);
ev_timer_set(&ps->draw_timer, 0, 0);
ev_timer_start(ps->loop, &ps->draw_timer);
} else {
// Let's draw our first frame!
queue_redraw(ps);

4
src/render.c
View file

@ -969,7 +969,7 @@ win_blur_background(session_t *ps, struct managed_win *w, xcb_render_picture_t t
/// paint all windows
/// region = ??
/// region_real = the damage region
void paint_all(session_t *ps, struct managed_win *t, bool ignore_damage) {
void paint_all(session_t *ps, struct managed_win *t) {
if (ps->o.xrender_sync_fence || (ps->drivers & DRIVER_NVIDIA)) {
if (ps->xsync_exists && !x_fence_sync(ps->c, ps->sync_fence)) {
log_error("x_fence_sync failed, xrender-sync-fence will be "
@ -984,7 +984,7 @@ void paint_all(session_t *ps, struct managed_win *t, bool ignore_damage) {
region_t region;
pixman_region32_init(&region);
int buffer_age = get_buffer_age(ps);
if (buffer_age == -1 || buffer_age > ps->ndamage || ignore_damage) {
if (buffer_age == -1 || buffer_age > ps->ndamage) {
pixman_region32_copy(&region, &ps->screen_reg);
} else {
for (int i = 0; i < get_buffer_age(ps); i++) {

2
src/render.h
View file

@ -37,7 +37,7 @@ void render(session_t *ps, int x, int y, int dx, int dy, int w, int h, int fullw
const glx_prog_main_t *pprogram, clip_t *clip);
void paint_one(session_t *ps, struct managed_win *w, const region_t *reg_paint);
void paint_all(session_t *ps, struct managed_win *const t, bool ignore_damage);
void paint_all(session_t *ps, struct managed_win *const t);
void free_picture(xcb_connection_t *c, xcb_render_picture_t *p);

100
src/statistics.c Normal file
View file

@ -0,0 +1,100 @@
//! Rendering statistics
//!
//! Tracks how long it takes to render a frame, for measuring performance, and for pacing
//! the frames.
#include "statistics.h"
#include "log.h"
#include "utils.h"
void render_statistics_init(struct render_statistics *rs, int window_size) {
*rs = (struct render_statistics){0};
rolling_window_init(&rs->render_times, window_size);
rolling_quantile_init_with_tolerance(&rs->render_time_quantile, window_size,
/* q */ 0.98, /* tolerance */ 0.01);
}
void render_statistics_add_vblank_time_sample(struct render_statistics *rs, int time_us) {
auto sample_sd = sqrt(cumulative_mean_and_var_get_var(&rs->vblank_time_us));
auto current_estimate = render_statistics_get_vblank_time(rs);
if (current_estimate != 0 && fabs((double)time_us - current_estimate) > sample_sd * 3) {
// Deviated from the mean by more than 3 sigma (p < 0.003)
log_debug("vblank time outlier: %d %f %f", time_us, rs->vblank_time_us.mean,
cumulative_mean_and_var_get_var(&rs->vblank_time_us));
// An outlier sample, this could mean things like refresh rate changes, so
// we reset the statistics. This could also be benign, but we like to be
// cautious.
cumulative_mean_and_var_init(&rs->vblank_time_us);
}
if (rs->vblank_time_us.mean != 0) {
auto nframes_in_10_seconds =
(unsigned int)(10. * 1000000. / rs->vblank_time_us.mean);
if (rs->vblank_time_us.n > 20 && rs->vblank_time_us.n > nframes_in_10_seconds) {
// We collected 10 seconds worth of samples, we assume the
// estimated refresh rate is stable. We will still reset the
// statistics if we get an outlier sample though, see above.
return;
}
}
cumulative_mean_and_var_update(&rs->vblank_time_us, time_us);
}
void render_statistics_add_render_time_sample(struct render_statistics *rs, int time_us) {
int oldest;
if (rolling_window_push_back(&rs->render_times, time_us, &oldest)) {
rolling_quantile_pop_front(&rs->render_time_quantile, oldest);
}
rolling_quantile_push_back(&rs->render_time_quantile, time_us);
}
/// How much time budget we should give to the backend for rendering, in microseconds.
///
/// A `divisor` is also returned, indicating the target framerate. The divisor is
/// the number of vblanks we should wait between each frame. A divisor of 1 means
/// full framerate, 2 means half framerate, etc.
unsigned int
render_statistics_get_budget(struct render_statistics *rs, unsigned int *divisor) {
if (rs->render_times.nelem < rs->render_times.window_size) {
// No valid render time estimates yet. Assume maximum budget.
*divisor = 1;
return UINT_MAX;
}
// N-th percentile of render times, see render_statistics_init for N.
auto render_time_percentile =
rolling_quantile_estimate(&rs->render_time_quantile, &rs->render_times);
auto vblank_time_us = render_statistics_get_vblank_time(rs);
if (vblank_time_us == 0) {
// We don't have a good estimate of the vblank time yet, so we
// assume we can finish in one vblank.
*divisor = 1;
} else {
*divisor =
(unsigned int)(render_time_percentile / rs->vblank_time_us.mean + 1);
}
return (unsigned int)render_time_percentile;
}
unsigned int render_statistics_get_vblank_time(struct render_statistics *rs) {
if (rs->vblank_time_us.n <= 20 || rs->vblank_time_us.mean < 100) {
// Not enough samples yet, or the vblank time is too short to be
// meaningful. Assume maximum budget.
return 0;
}
return (unsigned int)rs->vblank_time_us.mean;
}
void render_statistics_reset(struct render_statistics *rs) {
rolling_window_reset(&rs->render_times);
rolling_quantile_reset(&rs->render_time_quantile);
rs->vblank_time_us = (struct cumulative_mean_and_var){0};
}
void render_statistics_destroy(struct render_statistics *rs) {
render_statistics_reset(rs);
rolling_window_destroy(&rs->render_times);
rolling_quantile_destroy(&rs->render_time_quantile);
}

33
src/statistics.h Normal file
View file

@ -0,0 +1,33 @@
#pragma once
#include "utils.h"
#define NTIERS (3)
struct render_statistics {
/// Rolling window of rendering times (in us) and the tiers they belong to.
/// We keep track of the tiers because the vblank time estimate can change over
/// time.
struct rolling_window render_times;
/// Estimate the 95-th percentile of rendering times
struct rolling_quantile render_time_quantile;
/// Time between each vblanks
struct cumulative_mean_and_var vblank_time_us;
};
void render_statistics_init(struct render_statistics *rs, int window_size);
void render_statistics_reset(struct render_statistics *rs);
void render_statistics_destroy(struct render_statistics *rs);
void render_statistics_add_vblank_time_sample(struct render_statistics *rs, int time_us);
void render_statistics_add_render_time_sample(struct render_statistics *rs, int time_us);
/// How much time budget we should give to the backend for rendering, in microseconds.
///
/// A `divisor` is also returned, indicating the target framerate. The divisor is
/// the number of vblanks we should wait between each frame. A divisor of 1 means
/// full framerate, 2 means half framerate, etc.
unsigned int
render_statistics_get_budget(struct render_statistics *rs, unsigned int *divisor);
unsigned int render_statistics_get_vblank_time(struct render_statistics *rs);

233
src/utils.c
View file

@ -48,7 +48,43 @@ int next_power_of_two(int n) {
return n;
}
/// Track the rolling maximum of a stream of integers.
void rolling_window_destroy(struct rolling_window *rw) {
free(rw->elem);
rw->elem = NULL;
}
void rolling_window_reset(struct rolling_window *rw) {
rw->nelem = 0;
rw->elem_head = 0;
}
void rolling_window_init(struct rolling_window *rw, int size) {
rw->elem = ccalloc(size, int);
rw->window_size = size;
rolling_window_reset(rw);
}
int rolling_window_pop_front(struct rolling_window *rw) {
assert(rw->nelem > 0);
auto ret = rw->elem[rw->elem_head];
rw->elem_head = (rw->elem_head + 1) % rw->window_size;
rw->nelem--;
return ret;
}
bool rolling_window_push_back(struct rolling_window *rw, int val, int *front) {
bool full = rw->nelem == rw->window_size;
if (full) {
*front = rolling_window_pop_front(rw);
}
rw->elem[(rw->elem_head + rw->nelem) % rw->window_size] = val;
rw->nelem++;
return full;
}
/// Track the maximum member of a FIFO queue of integers. Integers are pushed to the back
/// and popped from the front, the maximum of the current members in the queue is
/// tracked.
struct rolling_max {
/// A priority queue holding the indices of the maximum element candidates.
/// The head of the queue is the index of the maximum element.
@ -59,32 +95,26 @@ struct rolling_max {
/// it's called the "original" indices.
int *p;
int p_head, np;
/// The elemets
int *elem;
int elem_head, nelem;
int window_size;
/// The maximum number of in flight elements.
int capacity;
};
void rolling_max_destroy(struct rolling_max *rm) {
free(rm->elem);
free(rm->p);
free(rm);
}
struct rolling_max *rolling_max_new(int size) {
struct rolling_max *rolling_max_new(int capacity) {
auto rm = ccalloc(1, struct rolling_max);
if (!rm) {
return NULL;
}
rm->p = ccalloc(size, int);
rm->elem = ccalloc(size, int);
rm->window_size = size;
if (!rm->p || !rm->elem) {
rm->p = ccalloc(capacity, int);
if (!rm->p) {
goto err;
}
rm->capacity = capacity;
return rm;
@ -96,33 +126,21 @@ err:
void rolling_max_reset(struct rolling_max *rm) {
rm->p_head = 0;
rm->np = 0;
rm->nelem = 0;
rm->elem_head = 0;
}
void rolling_max_push(struct rolling_max *rm, int val) {
#define IDX(n) ((n) % rm->window_size)
if (rm->nelem == rm->window_size) {
auto old_head = rm->elem_head;
// Discard the oldest element.
// rm->elem.pop_front();
rm->nelem--;
rm->elem_head = IDX(rm->elem_head + 1);
// Remove discarded element from the priority queue too.
assert(rm->np);
if (rm->p[rm->p_head] == old_head) {
// rm->p.pop_front()
rm->p_head = IDX(rm->p_head + 1);
rm->np--;
}
#define IDX(n) ((n) % rm->capacity)
/// Remove the oldest element in the window. The caller must maintain the list of elements
/// themselves, i.e. the behavior is undefined if `front` does not 1match the oldest
/// element.
void rolling_max_pop_front(struct rolling_max *rm, int front) {
if (rm->p[rm->p_head] == front) {
// rm->p.pop_front()
rm->p_head = IDX(rm->p_head + 1);
rm->np--;
}
}
// Add the new element to the queue.
// rm->elem.push_back(val)
rm->elem[IDX(rm->elem_head + rm->nelem)] = val;
rm->nelem++;
void rolling_max_push_back(struct rolling_max *rm, int val) {
// Update the prority queue.
// Remove all elements smaller than the new element from the queue. Because
// the new element will become the maximum element before them, and since they
@ -130,7 +148,7 @@ void rolling_max_push(struct rolling_max *rm, int val) {
// element, so they will never become the maximum element.
while (rm->np) {
int p_tail = IDX(rm->p_head + rm->np - 1);
if (rm->elem[rm->p[p_tail]] > val) {
if (rm->p[p_tail] > val) {
break;
}
// rm->p.pop_back()
@ -138,108 +156,119 @@ void rolling_max_push(struct rolling_max *rm, int val) {
}
// Add the new element to the end of the queue.
// rm->p.push_back(rm->start_index + rm->nelem - 1)
rm->p[IDX(rm->p_head + rm->np)] = IDX(rm->elem_head + rm->nelem - 1);
assert(rm->np < rm->capacity);
rm->p[IDX(rm->p_head + rm->np)] = val;
rm->np++;
#undef IDX
}
#undef IDX
int rolling_max_get_max(struct rolling_max *rm) {
if (rm->np == 0) {
return INT_MIN;
}
return rm->elem[rm->p[rm->p_head]];
return rm->p[rm->p_head];
}
TEST_CASE(rolling_max_test) {
#define NELEM 15
struct rolling_window queue;
rolling_window_init(&queue, 3);
auto rm = rolling_max_new(3);
const int data[NELEM] = {1, 2, 3, 1, 4, 5, 2, 3, 6, 5, 4, 3, 2, 0, 0};
const int expected_max[NELEM] = {1, 2, 3, 3, 4, 5, 5, 5, 6, 6, 6, 5, 4, 3, 2};
int max[NELEM] = {0};
for (int i = 0; i < NELEM; i++) {
rolling_max_push(rm, data[i]);
int front;
bool full = rolling_window_push_back(&queue, data[i], &front);
if (full) {
rolling_max_pop_front(rm, front);
}
rolling_max_push_back(rm, data[i]);
max[i] = rolling_max_get_max(rm);
}
rolling_window_destroy(&queue);
rolling_max_destroy(rm);
TEST_TRUE(memcmp(max, expected_max, sizeof(max)) == 0);
#undef NELEM
}
/// A rolling average of a stream of integers.
struct rolling_avg {
/// The sum of the elements in the window.
int64_t sum;
// Find the k-th smallest element in an array.
int quickselect(int *elems, int nelem, int k) {
int l = 0, r = nelem; // [l, r) is the range of candidates
while (l != r) {
int pivot = elems[l];
int i = l, j = r;
while (i < j) {
while (i < j && elems[--j] >= pivot) {
}
elems[i] = elems[j];
while (i < j && elems[++i] <= pivot) {
}
elems[j] = elems[i];
}
elems[i] = pivot;
/// The elements in the window.
int *elem;
int head, nelem;
if (i == k) {
break;
}
int window_size;
};
struct rolling_avg *rolling_avg_new(int size) {
auto rm = ccalloc(1, struct rolling_avg);
if (!rm) {
return NULL;
if (i < k) {
l = i + 1;
} else {
r = i;
}
}
rm->elem = ccalloc(size, int);
rm->window_size = size;
if (!rm->elem) {
free(rm);
return NULL;
}
return rm;
return elems[k];
}
void rolling_avg_destroy(struct rolling_avg *rm) {
free(rm->elem);
free(rm);
void rolling_quantile_init(struct rolling_quantile *rq, int capacity, int mink, int maxk) {
*rq = (struct rolling_quantile){0};
rq->tmp_buffer = malloc(sizeof(int) * (size_t)capacity);
rq->capacity = capacity;
rq->min_target_rank = mink;
rq->max_target_rank = maxk;
}
void rolling_avg_reset(struct rolling_avg *ra) {
ra->sum = 0;
ra->nelem = 0;
ra->head = 0;
void rolling_quantile_init_with_tolerance(struct rolling_quantile *rq, int window_size,
double target, double tolerance) {
rolling_quantile_init(rq, window_size, (int)((target - tolerance) * window_size),
(int)((target + tolerance) * window_size));
}
void rolling_avg_push(struct rolling_avg *ra, int val) {
if (ra->nelem == ra->window_size) {
// Discard the oldest element.
// rm->elem.pop_front();
ra->sum -= ra->elem[ra->head % ra->window_size];
ra->nelem--;
ra->head = (ra->head + 1) % ra->window_size;
void rolling_quantile_reset(struct rolling_quantile *rq) {
rq->current_rank = 0;
rq->estimate = 0;
}
void rolling_quantile_destroy(struct rolling_quantile *rq) {
free(rq->tmp_buffer);
}
int rolling_quantile_estimate(struct rolling_quantile *rq, struct rolling_window *elements) {
if (rq->current_rank < rq->min_target_rank || rq->current_rank > rq->max_target_rank) {
if (elements->nelem != elements->window_size) {
return INT_MIN;
}
// Re-estimate the quantile.
assert(elements->nelem <= rq->capacity);
rolling_window_copy_to_array(elements, rq->tmp_buffer);
const int target_rank =
rq->min_target_rank + (rq->max_target_rank - rq->min_target_rank) / 2;
rq->estimate = quickselect(rq->tmp_buffer, elements->nelem, target_rank);
rq->current_rank = target_rank;
}
// Add the new element to the queue.
// rm->elem.push_back(val)
ra->elem[(ra->head + ra->nelem) % ra->window_size] = val;
ra->sum += val;
ra->nelem++;
return rq->estimate;
}
double rolling_avg_get_avg(struct rolling_avg *ra) {
if (ra->nelem == 0) {
return 0;
void rolling_quantile_push_back(struct rolling_quantile *rq, int x) {
if (x <= rq->estimate) {
rq->current_rank++;
}
return (double)ra->sum / (double)ra->nelem;
}
TEST_CASE(rolling_avg_test) {
#define NELEM 15
auto rm = rolling_avg_new(3);
const int data[NELEM] = {1, 2, 3, 1, 4, 5, 2, 3, 6, 5, 4, 3, 2, 0, 0};
const double expected_avg[NELEM] = {
1, 1.5, 2, 2, 8.0 / 3.0, 10.0 / 3.0, 11.0 / 3.0, 10.0 / 3.0,
11.0 / 3.0, 14.0 / 3.0, 5, 4, 3, 5.0 / 3.0, 2.0 / 3.0};
double avg[NELEM] = {0};
for (int i = 0; i < NELEM; i++) {
rolling_avg_push(rm, data[i]);
avg[i] = rolling_avg_get_avg(rm);
}
for (int i = 0; i < NELEM; i++) {
TEST_EQUAL(avg[i], expected_avg[i]);
void rolling_quantile_pop_front(struct rolling_quantile *rq, int x) {
if (x <= rq->estimate) {
rq->current_rank--;
}
}

96
src/utils.h
View file

@ -17,6 +17,7 @@
#include <time.h>
#include "compiler.h"
#include "log.h"
#include "types.h"
#define ARR_SIZE(arr) (sizeof(arr) / sizeof(arr[0]))
@ -289,20 +290,97 @@ static inline void free_charpp(char **str) {
///
int next_power_of_two(int n);
struct rolling_window {
int *elem;
int elem_head, nelem;
int window_size;
};
void rolling_window_destroy(struct rolling_window *rw);
void rolling_window_reset(struct rolling_window *rw);
void rolling_window_init(struct rolling_window *rw, int size);
int rolling_window_pop_front(struct rolling_window *rw);
bool rolling_window_push_back(struct rolling_window *rw, int val, int *front);
/// Copy the contents of the rolling window to an array. The array is assumed to
/// have enough space to hold the contents of the rolling window.
static inline void attr_unused rolling_window_copy_to_array(struct rolling_window *rw,
int *arr) {
// The length from head to the end of the array
auto head_len = (size_t)(rw->window_size - rw->elem_head);
if (head_len >= (size_t)rw->nelem) {
memcpy(arr, rw->elem + rw->elem_head, sizeof(int) * (size_t)rw->nelem);
} else {
auto tail_len = (size_t)((size_t)rw->nelem - head_len);
memcpy(arr, rw->elem + rw->elem_head, sizeof(int) * head_len);
memcpy(arr + head_len, rw->elem, sizeof(int) * tail_len);
}
}
struct rolling_max;
struct rolling_max *rolling_max_new(int window_size);
void rolling_max_free(struct rolling_max *rm);
struct rolling_max *rolling_max_new(int capacity);
void rolling_max_destroy(struct rolling_max *rm);
void rolling_max_reset(struct rolling_max *rm);
void rolling_max_push(struct rolling_max *rm, int val);
void rolling_max_pop_front(struct rolling_max *rm, int front);
void rolling_max_push_back(struct rolling_max *rm, int val);
int rolling_max_get_max(struct rolling_max *rm);
struct rolling_avg;
struct rolling_avg *rolling_avg_new(int window_size);
void rolling_avg_free(struct rolling_avg *ra);
void rolling_avg_reset(struct rolling_avg *ra);
void rolling_avg_push(struct rolling_avg *ra, int val);
double rolling_avg_get_avg(struct rolling_avg *ra);
/// Estimate the mean and variance of random variable X using Welford's online
/// algorithm.
struct cumulative_mean_and_var {
double mean;
double m2;
unsigned int n;
};
static inline attr_unused void
cumulative_mean_and_var_init(struct cumulative_mean_and_var *cmv) {
*cmv = (struct cumulative_mean_and_var){0};
}
static inline attr_unused void
cumulative_mean_and_var_update(struct cumulative_mean_and_var *cmv, double x) {
if (cmv->n == UINT_MAX) {
// We have too many elements, let's keep the mean and variance.
return;
}
cmv->n++;
double delta = x - cmv->mean;
cmv->mean += delta / (double)cmv->n;
cmv->m2 += delta * (x - cmv->mean);
}
static inline attr_unused double
cumulative_mean_and_var_get_var(struct cumulative_mean_and_var *cmv) {
if (cmv->n < 2) {
return 0;
}
return cmv->m2 / (double)(cmv->n - 1);
}
// Find the k-th smallest element in an array.
int quickselect(int *elems, int nelem, int k);
/// A naive quantile estimator.
///
/// Estimates the N-th percentile of a random variable X in a sliding window.
struct rolling_quantile {
int current_rank;
int min_target_rank, max_target_rank;
int estimate;
int capacity;
int *tmp_buffer;
};
void rolling_quantile_init(struct rolling_quantile *rq, int capacity, int mink, int maxk);
void rolling_quantile_init_with_tolerance(struct rolling_quantile *rq, int window_size,
double target, double tolerance);
void rolling_quantile_reset(struct rolling_quantile *rq);
void rolling_quantile_destroy(struct rolling_quantile *rq);
int rolling_quantile_estimate(struct rolling_quantile *rq, struct rolling_window *elements);
void rolling_quantile_push_back(struct rolling_quantile *rq, int x);
void rolling_quantile_pop_front(struct rolling_quantile *rq, int x);
// Some versions of the Android libc do not have timespec_get(), use
// clock_gettime() instead.