From: Peter Lieven <pl@kamp.de>
To: qemu-devel@nongnu.org
Cc: sw@weilnetz.de, Peter Lieven <pl@kamp.de>,
xiawenc@linux.vnet.ibm.com, aliguori@amazon.com
Subject: [Qemu-devel] [PATCHv4 3/6] ui/vnc: optimize dirty bitmap tracking
Date: Wed, 8 Jan 2014 10:08:35 +0100 [thread overview]
Message-ID: <1389172118-25402-4-git-send-email-pl@kamp.de> (raw)
In-Reply-To: <1389172118-25402-1-git-send-email-pl@kamp.de>
vnc_update_client currently scans the dirty bitmap of each client
bitwise which is a very costly operation if only few bits are dirty.
vnc_refresh_server_surface does almost the same.
this patch optimizes both by utilizing the heavily optimized
function find_next_bit to find the offset of the next dirty
bit in the dirty bitmaps.
The following artifical test (just the bitmap operation part) running
vnc_update_client 65536 times on a 2560x2048 surface illustrates the
performance difference:
All bits clean - vnc_update_client_new: 0.07 secs
vnc_update_client_old: 10.98 secs
All bits dirty - vnc_update_client_new: 11.26 secs
vnc_update_client_old: 20.19 secs
Few bits dirty - vnc_update_client_new: 0.08 secs
vnc_update_client_old: 10.98 secs
The case for all bits dirty is still rather slow, this
is due to the implementation of find_and_clear_dirty_height.
This will be addresses in a separate patch.
Signed-off-by: Peter Lieven <pl@kamp.de>
---
ui/vnc.c | 155 ++++++++++++++++++++++++++++++++++----------------------------
ui/vnc.h | 4 ++
2 files changed, 88 insertions(+), 71 deletions(-)
diff --git a/ui/vnc.c b/ui/vnc.c
index f42398d..3412cdf 100644
--- a/ui/vnc.c
+++ b/ui/vnc.c
@@ -572,6 +572,15 @@ void *vnc_server_fb_ptr(VncDisplay *vd, int x, int y)
ptr += x * VNC_SERVER_FB_BYTES;
return ptr;
}
+/* this sets only the visible pixels of a dirty bitmap */
+#define VNC_SET_VISIBLE_PIXELS_DIRTY(bitmap, w, h) {\
+ int y;\
+ memset(bitmap, 0x00, sizeof(bitmap));\
+ for (y = 0; y < h; y++) {\
+ bitmap_set(bitmap[y], 0,\
+ DIV_ROUND_UP(w, VNC_DIRTY_PIXELS_PER_BIT));\
+ } \
+ }
static void vnc_dpy_switch(DisplayChangeListener *dcl,
DisplaySurface *surface)
@@ -597,7 +606,9 @@ static void vnc_dpy_switch(DisplayChangeListener *dcl,
qemu_pixman_image_unref(vd->guest.fb);
vd->guest.fb = pixman_image_ref(surface->image);
vd->guest.format = surface->format;
- memset(vd->guest.dirty, 0xFF, sizeof(vd->guest.dirty));
+ VNC_SET_VISIBLE_PIXELS_DIRTY(vd->guest.dirty,
+ surface_width(vd->ds),
+ surface_height(vd->ds));
QTAILQ_FOREACH(vs, &vd->clients, next) {
vnc_colordepth(vs);
@@ -605,7 +616,9 @@ static void vnc_dpy_switch(DisplayChangeListener *dcl,
if (vs->vd->cursor) {
vnc_cursor_define(vs);
}
- memset(vs->dirty, 0xFF, sizeof(vs->dirty));
+ VNC_SET_VISIBLE_PIXELS_DIRTY(vs->dirty,
+ surface_width(vd->ds),
+ surface_height(vd->ds));
}
}
@@ -891,10 +904,9 @@ static int vnc_update_client(VncState *vs, int has_dirty)
VncDisplay *vd = vs->vd;
VncJob *job;
int y;
- int width, height;
+ int height;
int n = 0;
-
if (vs->output.offset && !vs->audio_cap && !vs->force_update)
/* kernel send buffers are full -> drop frames to throttle */
return 0;
@@ -910,39 +922,27 @@ static int vnc_update_client(VncState *vs, int has_dirty)
*/
job = vnc_job_new(vs);
- width = MIN(pixman_image_get_width(vd->server), vs->client_width);
height = MIN(pixman_image_get_height(vd->server), vs->client_height);
- for (y = 0; y < height; y++) {
- int x;
- int last_x = -1;
- for (x = 0; x < width / VNC_DIRTY_PIXELS_PER_BIT; x++) {
- if (test_and_clear_bit(x, vs->dirty[y])) {
- if (last_x == -1) {
- last_x = x;
- }
- } else {
- if (last_x != -1) {
- int h = find_and_clear_dirty_height(vs, y, last_x, x,
- height);
-
- n += vnc_job_add_rect(job,
- last_x * VNC_DIRTY_PIXELS_PER_BIT,
- y,
- (x - last_x) *
- VNC_DIRTY_PIXELS_PER_BIT,
- h);
- }
- last_x = -1;
- }
- }
- if (last_x != -1) {
- int h = find_and_clear_dirty_height(vs, y, last_x, x, height);
- n += vnc_job_add_rect(job, last_x * VNC_DIRTY_PIXELS_PER_BIT,
- y,
- (x - last_x) * VNC_DIRTY_PIXELS_PER_BIT,
- h);
+ y = 0;
+ for (;;) {
+ int x, h;
+ unsigned long x2;
+ unsigned long offset = find_next_bit((unsigned long *) &vs->dirty,
+ height * VNC_DIRTY_BPL(vs),
+ y * VNC_DIRTY_BPL(vs));
+ if (offset == height * VNC_DIRTY_BPL(vs)) {
+ /* no more dirty bits */
+ break;
}
+ y = offset / VNC_DIRTY_BPL(vs);
+ x = offset % VNC_DIRTY_BPL(vs);
+ x2 = find_next_zero_bit((unsigned long *) &vs->dirty[y],
+ VNC_DIRTY_BPL(vs), x);
+ bitmap_clear(vs->dirty[y], x, x2 - x);
+ h = find_and_clear_dirty_height(vs, y, x, x2, height);
+ n += vnc_job_add_rect(job, x * VNC_DIRTY_PIXELS_PER_BIT, y,
+ (x2 - x) * VNC_DIRTY_PIXELS_PER_BIT, h);
}
vnc_job_push(job);
@@ -2680,8 +2680,8 @@ static int vnc_refresh_server_surface(VncDisplay *vd)
int width = pixman_image_get_width(vd->guest.fb);
int height = pixman_image_get_height(vd->guest.fb);
int y;
- uint8_t *guest_row;
- uint8_t *server_row;
+ uint8_t *guest_row0 = NULL, *server_row0;
+ int guest_stride = 0, server_stride;
int cmp_bytes;
VncState *vs;
int has_dirty = 0;
@@ -2706,44 +2706,57 @@ static int vnc_refresh_server_surface(VncDisplay *vd)
if (vd->guest.format != VNC_SERVER_FB_FORMAT) {
int width = pixman_image_get_width(vd->server);
tmpbuf = qemu_pixman_linebuf_create(VNC_SERVER_FB_FORMAT, width);
- }
- guest_row = (uint8_t *)pixman_image_get_data(vd->guest.fb);
- server_row = (uint8_t *)pixman_image_get_data(vd->server);
- for (y = 0; y < height; y++) {
- if (!bitmap_empty(vd->guest.dirty[y], VNC_DIRTY_BITS)) {
- int x;
- uint8_t *guest_ptr;
- uint8_t *server_ptr;
-
- if (vd->guest.format != VNC_SERVER_FB_FORMAT) {
- qemu_pixman_linebuf_fill(tmpbuf, vd->guest.fb, width, 0, y);
- guest_ptr = (uint8_t *)pixman_image_get_data(tmpbuf);
- } else {
- guest_ptr = guest_row;
- }
- server_ptr = server_row;
+ } else {
+ guest_row0 = (uint8_t *)pixman_image_get_data(vd->guest.fb);
+ guest_stride = pixman_image_get_stride(vd->guest.fb);
+ }
+ server_row0 = (uint8_t *)pixman_image_get_data(vd->server);
+ server_stride = pixman_image_get_stride(vd->server);
+
+ y = 0;
+ for (;;) {
+ int x;
+ uint8_t *guest_ptr, *server_ptr;
+ unsigned long offset = find_next_bit((unsigned long *) &vd->guest.dirty,
+ height * VNC_DIRTY_BPL(&vd->guest),
+ y * VNC_DIRTY_BPL(&vd->guest));
+ if (offset == height * VNC_DIRTY_BPL(&vd->guest)) {
+ /* no more dirty bits */
+ break;
+ }
+ y = offset / VNC_DIRTY_BPL(&vd->guest);
+ x = offset % VNC_DIRTY_BPL(&vd->guest);
- for (x = 0; x + VNC_DIRTY_PIXELS_PER_BIT - 1 < width;
- x += VNC_DIRTY_PIXELS_PER_BIT, guest_ptr += cmp_bytes,
- server_ptr += cmp_bytes) {
- if (!test_and_clear_bit((x / VNC_DIRTY_PIXELS_PER_BIT),
- vd->guest.dirty[y])) {
- continue;
- }
- if (memcmp(server_ptr, guest_ptr, cmp_bytes) == 0) {
- continue;
- }
- memcpy(server_ptr, guest_ptr, cmp_bytes);
- if (!vd->non_adaptive)
- vnc_rect_updated(vd, x, y, &tv);
- QTAILQ_FOREACH(vs, &vd->clients, next) {
- set_bit((x / VNC_DIRTY_PIXELS_PER_BIT), vs->dirty[y]);
- }
- has_dirty++;
+ server_ptr = server_row0 + y * server_stride + x * cmp_bytes;
+
+ if (vd->guest.format != VNC_SERVER_FB_FORMAT) {
+ qemu_pixman_linebuf_fill(tmpbuf, vd->guest.fb, width, 0, y);
+ guest_ptr = (uint8_t *)pixman_image_get_data(tmpbuf);
+ } else {
+ guest_ptr = guest_row0 + y * guest_stride;
+ }
+ guest_ptr += x * cmp_bytes;
+
+ for (; x < DIV_ROUND_UP(width, VNC_DIRTY_PIXELS_PER_BIT);
+ x++, guest_ptr += cmp_bytes, server_ptr += cmp_bytes) {
+ if (!test_and_clear_bit(x, vd->guest.dirty[y])) {
+ continue;
+ }
+ if (memcmp(server_ptr, guest_ptr, cmp_bytes) == 0) {
+ continue;
+ }
+ memcpy(server_ptr, guest_ptr, cmp_bytes);
+ if (!vd->non_adaptive) {
+ vnc_rect_updated(vd, x * VNC_DIRTY_PIXELS_PER_BIT,
+ y, &tv);
}
+ QTAILQ_FOREACH(vs, &vd->clients, next) {
+ set_bit(x, vs->dirty[y]);
+ }
+ has_dirty++;
}
- guest_row += pixman_image_get_stride(vd->guest.fb);
- server_row += pixman_image_get_stride(vd->server);
+
+ y++;
}
qemu_pixman_image_unref(tmpbuf);
return has_dirty;
diff --git a/ui/vnc.h b/ui/vnc.h
index 4a8f33c..07e1f59 100644
--- a/ui/vnc.h
+++ b/ui/vnc.h
@@ -88,6 +88,10 @@ typedef void VncSendHextileTile(VncState *vs,
/* VNC_DIRTY_BITS is the number of bits in the dirty bitmap. */
#define VNC_DIRTY_BITS (VNC_MAX_WIDTH / VNC_DIRTY_PIXELS_PER_BIT)
+/* VNC_DIRTY_BPL (BPL = bits per line) might be greater than
+ * VNC_DIRTY_BITS due to alignment */
+#define VNC_DIRTY_BPL(x) (sizeof((x)->dirty) / VNC_MAX_HEIGHT * BITS_PER_BYTE)
+
#define VNC_STAT_RECT 64
#define VNC_STAT_COLS (VNC_MAX_WIDTH / VNC_STAT_RECT)
#define VNC_STAT_ROWS (VNC_MAX_HEIGHT / VNC_STAT_RECT)
--
1.7.9.5
next prev parent reply other threads:[~2014-01-08 9:07 UTC|newest]
Thread overview: 19+ messages / expand[flat|nested] mbox.gz Atom feed top
2014-01-08 9:08 [Qemu-devel] [PATCHv4 0/6] ui/vnc: update optimizations Peter Lieven
2014-01-08 9:08 ` [Qemu-devel] [PATCHv4 1/6] ui/vnc: introduce VNC_DIRTY_PIXELS_PER_BIT macro Peter Lieven
2014-01-09 4:44 ` Wenchao Xia
2014-01-08 9:08 ` [Qemu-devel] [PATCHv4 2/6] ui/vnc: derive cmp_bytes from VNC_DIRTY_PIXELS_PER_BIT Peter Lieven
2014-01-08 9:08 ` Peter Lieven [this message]
2014-01-09 7:37 ` [Qemu-devel] [PATCHv4 3/6] ui/vnc: optimize dirty bitmap tracking Wenchao Xia
2014-01-08 9:08 ` [Qemu-devel] [PATCHv4 4/6] ui/vnc: optimize clearing in find_and_clear_dirty_height() Peter Lieven
2014-01-09 7:58 ` Wenchao Xia
2014-01-08 9:08 ` [Qemu-devel] [PATCHv4 5/6] ui/vnc: optimize setting in vnc_dpy_update() Peter Lieven
2014-01-09 8:16 ` Wenchao Xia
2014-01-08 9:08 ` [Qemu-devel] [PATCHv4 6/6] ui/vnc: disable adaptive update calculations if not needed Peter Lieven
2014-01-09 8:29 ` Wenchao Xia
2014-01-09 16:25 ` Peter Lieven
2014-01-10 3:09 ` Wenchao Xia
2014-01-10 22:28 ` Peter Lieven
2014-01-13 2:42 ` Wenchao Xia
2014-01-13 8:27 ` Peter Lieven
2014-01-20 9:54 ` [Qemu-devel] [PATCHv4 0/6] ui/vnc: update optimizations Peter Lieven
2014-02-03 9:28 ` Peter Lieven
Reply instructions:
You may reply publicly to this message via plain-text email
using any one of the following methods:
* Save the following mbox file, import it into your mail client,
and reply-to-all from there: mbox
Avoid top-posting and favor interleaved quoting:
https://en.wikipedia.org/wiki/Posting_style#Interleaved_style
* Reply using the --to, --cc, and --in-reply-to
switches of git-send-email(1):
git send-email \
--in-reply-to=1389172118-25402-4-git-send-email-pl@kamp.de \
--to=pl@kamp.de \
--cc=aliguori@amazon.com \
--cc=qemu-devel@nongnu.org \
--cc=sw@weilnetz.de \
--cc=xiawenc@linux.vnet.ibm.com \
/path/to/YOUR_REPLY
https://kernel.org/pub/software/scm/git/docs/git-send-email.html
* If your mail client supports setting the In-Reply-To header
via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line
before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).