[PATCH 0/6] Add video damage tracking
This patch set speeds up graphics output on ARM by a factor of 60x.
On most ARM SBCs, we keep the frame buffer in DRAM and map it as cached, but need it accessible by the display controller which reads directly from a later point of consistency. Hence, we flush the frame buffer to DRAM on every change. The full frame buffer.
Unfortunately, with the advent of 4k displays, we are seeing frame buffers that can take a while to flush out. This was reported by Da Xue with grub, which happily print 1000s of spaces on the screen to draw a menu. Every printed space triggers a cache flush.
This patch set implements the easiest mitigation against this problem: Damage tracking. We remember the lowest common denominator region that was touched since the last video_sync() call and only flush that.
With this patch set applied, we reduce drawing a large grub menu (with serial console attached for size information) on an RK3399-ROC system at 1440p from 55 seconds to less than 1 second.
Alternatives considered:
1) Lazy sync - Sandbox does this. It only calls video_sync(true) ever so often. We are missing timers to do this generically.
2) Double buffering - We could try to identify whether anything changed at all and only draw to the FB if it did. That would require maintaining a second buffer that we need to scan.
3) Text buffer - Maintain a buffer of all text printed on the screen with respective location. Don't write if the old and new character are identical. This would limit applicability to text only and is an optimization on top of this patch set.
4) Hash screen lines - Create a hash (sha256?) over every line when it changes. Only flush when it does. I'm not sure if this would waste more time, memory and cache than the current approach. It would make full screen updates much more expensive.
Alexander Graf (6): dm: video: Add damage tracking API dm: video: Add damage notification on display clear vidconsole: Add damage notifications to all vidconsole drivers video: Add damage notification on bmp display efi_loader: GOP: Add damage notification on BLT video: Only dcache flush damaged lines
drivers/video/Kconfig | 15 ++++++ drivers/video/console_normal.c | 10 ++++ drivers/video/console_rotate.c | 18 +++++++ drivers/video/console_truetype.c | 12 +++++ drivers/video/video-uclass.c | 89 +++++++++++++++++++++++++++++--- drivers/video/video_bmp.c | 2 + include/video.h | 39 +++++++++++++- lib/efi_loader/efi_gop.c | 11 ++++ 8 files changed, 187 insertions(+), 9 deletions(-)
We are going to introduce image damage tracking to fasten up screen refresh on large displays. This patch adds damage tracking for up to one rectangle of the screen which is typically enough to hold blt or text print updates. Callers into this API and a reduced dcache flush code path will follow in later patches.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer --- drivers/video/Kconfig | 15 ++++++++++++++ drivers/video/video-uclass.c | 40 ++++++++++++++++++++++++++++++++++++ include/video.h | 39 +++++++++++++++++++++++++++++++++-- 3 files changed, 92 insertions(+), 2 deletions(-)
diff --git a/drivers/video/Kconfig b/drivers/video/Kconfig index 965b587927..9e1c409b37 100644 --- a/drivers/video/Kconfig +++ b/drivers/video/Kconfig @@ -64,6 +64,21 @@ config VIDEO_COPY To use this, your video driver must set @copy_base in struct video_uc_plat.
+config VIDEO_DAMAGE + bool "Enable damage tracking of frame buffer regions" + depends on DM_VIDEO + default y if ARM && !SYS_DCACHE_OFF + help + On some machines (most ARM), the display frame buffer resides in + RAM. To make the display controller pick up screen updates, we + have to flush frame buffer contents from CPU caches into RAM which + can be a slow operation. + + This patch adds damage tracking to collect information about regions + that received updates. When we want to sync, we then only flush + regions of the frame buffer that were modified before, speeding up + screen refreshes significantly. + config BACKLIGHT_PWM bool "Generic PWM based Backlight Driver" depends on BACKLIGHT && DM_PWM diff --git a/drivers/video/video-uclass.c b/drivers/video/video-uclass.c index 01e8af5ac6..496aa56843 100644 --- a/drivers/video/video-uclass.c +++ b/drivers/video/video-uclass.c @@ -21,6 +21,8 @@ #include <dm/device_compat.h> #include <dm/device-internal.h> #include <dm/uclass-internal.h> +#include <linux/types.h> +#include <linux/bitmap.h> #ifdef CONFIG_SANDBOX #include <asm/sdl.h> #endif @@ -180,6 +182,44 @@ void video_set_default_colors(struct udevice *dev, bool invert) priv->colour_bg = vid_console_color(priv, back); }
+#ifdef CONFIG_VIDEO_DAMAGE +/* Notify about changes in the frame buffer */ +int video_damage(struct udevice *vid, int x, int y, int width, int height) +{ + struct video_priv *priv = dev_get_uclass_priv(vid); + int endx = x + width; + int endy = y + height; + + if (x > priv->xsize) + return 0; + + if (y > priv->ysize) + return 0; + + if (endx > priv->xsize) + endx = priv->xsize; + + if (endy > priv->ysize) + endy = priv->ysize; + + if (priv->damage.endx && priv->damage.endy) { + /* Span a rectangle across all old and new damage */ + priv->damage.x = min(x, priv->damage.x); + priv->damage.y = min(y, priv->damage.y); + priv->damage.endx = max(endx, priv->damage.endx); + priv->damage.endy = max(endy, priv->damage.endy); + } else { + /* First damage, setting the rectangle to span it */ + priv->damage.x = x; + priv->damage.y = y; + priv->damage.endx = endx; + priv->damage.endy = endy; + } + + return 0; +} +#endif + /* Flush video activity to the caches */ int video_sync(struct udevice *vid, bool force) { diff --git a/include/video.h b/include/video.h index 43e2c89977..98592eb19a 100644 --- a/include/video.h +++ b/include/video.h @@ -109,6 +109,14 @@ struct video_priv { void *fb; int fb_size; void *copy_fb; +#ifdef CONFIG_VIDEO_DAMAGE + struct { + int x; + int y; + int endx; + int endy; + } damage; +#endif int line_length; u32 colour_fg; u32 colour_bg; @@ -167,8 +175,9 @@ int video_clear(struct udevice *dev); * @return: 0 on success, error code otherwise * * Some frame buffers are cached or have a secondary frame buffer. This - * function syncs these up so that the current contents of the U-Boot frame - * buffer are displayed to the user. + * function syncs the damaged parts of them up so that the current contents + * of the U-Boot frame buffer are displayed to the user. It clears the damage + * buffer. */ int video_sync(struct udevice *vid, bool force);
@@ -268,6 +277,32 @@ static inline int video_sync_copy_all(struct udevice *dev)
#endif
+#ifdef CONFIG_VIDEO_DAMAGE +/** + * video_damage() - Notify the video subsystem about screen updates. + * + * @vid: Device to sync + * @x: Upper left X coordinate of the damaged rectangle + * @y: Upper left Y coordinate of the damaged rectangle + * @width: Width of the damaged rectangle + * @height: Height of the damaged rectangle + * + * @return: 0 + * + * Some frame buffers are cached or have a secondary frame buffer. This + * function notifies the video subsystem about rectangles that were updated + * within the frame buffer. They may only get written to the screen on the + * next call to video_sync(). + */ +int video_damage(struct udevice *vid, int x, int y, int width, int height); +#else +static inline int video_damage(struct udevice *vid, int x, int y, int width, + int height) +{ + return 0; +} +#endif /* CONFIG_VIDEO_DAMAGE */ + /** * video_is_active() - Test if one video device it active *
Let's report the video damage when we clear the screen. This way we can later lazily flush only relevant regions to hardware.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer --- drivers/video/video-uclass.c | 2 ++ 1 file changed, 2 insertions(+)
diff --git a/drivers/video/video-uclass.c b/drivers/video/video-uclass.c index 496aa56843..9ac1974670 100644 --- a/drivers/video/video-uclass.c +++ b/drivers/video/video-uclass.c @@ -153,6 +153,8 @@ int video_clear(struct udevice *dev) if (ret) return ret;
+ video_damage(dev, 0, 0, priv->xsize, priv->ysize); + return video_sync(dev, false); }
Now that we have a damage tracking API, let's populate damage done by vidconsole drivers. We try to declare as little memory as damaged as possible, with the exception of rotated screens that I couldn't get my head wrapped around. On those, we revert to the old behavior and mark the full screen as damaged on every update.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer --- drivers/video/console_normal.c | 10 ++++++++++ drivers/video/console_rotate.c | 18 ++++++++++++++++++ drivers/video/console_truetype.c | 12 ++++++++++++ 3 files changed, 40 insertions(+)
diff --git a/drivers/video/console_normal.c b/drivers/video/console_normal.c index 04f022491e..5b5586fd3e 100644 --- a/drivers/video/console_normal.c +++ b/drivers/video/console_normal.c @@ -57,6 +57,9 @@ static int console_normal_set_row(struct udevice *dev, uint row, int clr) if (ret) return ret;
+ video_damage(dev->parent, 0, VIDEO_FONT_HEIGHT * row, vid_priv->xsize, + VIDEO_FONT_HEIGHT); + return 0; }
@@ -76,6 +79,9 @@ static int console_normal_move_rows(struct udevice *dev, uint rowdst, if (ret) return ret;
+ video_damage(dev->parent, 0, VIDEO_FONT_HEIGHT * rowdst, vid_priv->xsize, + VIDEO_FONT_HEIGHT * count); + return 0; }
@@ -143,6 +149,10 @@ static int console_normal_putc_xy(struct udevice *dev, uint x_frac, uint y, } line += vid_priv->line_length; } + + video_damage(dev->parent, VID_TO_PIXEL(x_frac), y, VIDEO_FONT_WIDTH, + VIDEO_FONT_HEIGHT); + ret = vidconsole_sync_copy(dev, start, line); if (ret) return ret; diff --git a/drivers/video/console_rotate.c b/drivers/video/console_rotate.c index 36c8d0609d..4d5084e8d1 100644 --- a/drivers/video/console_rotate.c +++ b/drivers/video/console_rotate.c @@ -57,6 +57,8 @@ static int console_set_row_1(struct udevice *dev, uint row, int clr) if (ret) return ret;
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return 0; }
@@ -83,6 +85,8 @@ static int console_move_rows_1(struct udevice *dev, uint rowdst, uint rowsrc, dst += vid_priv->line_length; }
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return 0; }
@@ -150,6 +154,8 @@ static int console_putc_xy_1(struct udevice *dev, uint x_frac, uint y, char ch) if (ret) return ret;
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return VID_TO_POS(VIDEO_FONT_WIDTH); }
@@ -199,6 +205,8 @@ static int console_set_row_2(struct udevice *dev, uint row, int clr) if (ret) return ret;
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return 0; }
@@ -218,6 +226,8 @@ static int console_move_rows_2(struct udevice *dev, uint rowdst, uint rowsrc, vidconsole_memmove(dev, dst, src, VIDEO_FONT_HEIGHT * vid_priv->line_length * count);
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return 0; }
@@ -288,6 +298,8 @@ static int console_putc_xy_2(struct udevice *dev, uint x_frac, uint y, char ch) if (ret) return ret;
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return VID_TO_POS(VIDEO_FONT_WIDTH); }
@@ -335,6 +347,8 @@ static int console_set_row_3(struct udevice *dev, uint row, int clr) if (ret) return ret;
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return 0; }
@@ -359,6 +373,8 @@ static int console_move_rows_3(struct udevice *dev, uint rowdst, uint rowsrc, dst += vid_priv->line_length; }
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return 0; }
@@ -424,6 +440,8 @@ static int console_putc_xy_3(struct udevice *dev, uint x_frac, uint y, char ch) if (ret) return ret;
+ video_damage(dev->parent, 0, 0, vid_priv->xsize, vid_priv->ysize); + return VID_TO_POS(VIDEO_FONT_WIDTH); }
diff --git a/drivers/video/console_truetype.c b/drivers/video/console_truetype.c index c04b449a6d..8fab28fd15 100644 --- a/drivers/video/console_truetype.c +++ b/drivers/video/console_truetype.c @@ -168,6 +168,9 @@ static int console_truetype_set_row(struct udevice *dev, uint row, int clr) if (ret) return ret;
+ video_damage(dev->parent, 0, priv->y_charsize * rowdst, vid_priv->xsize, + priv->y_charsize); + return 0; }
@@ -192,6 +195,9 @@ static int console_truetype_move_rows(struct udevice *dev, uint rowdst, for (i = 0; i < priv->pos_ptr; i++) priv->pos[i].ypos -= diff;
+ video_damage(dev->parent, 0, priv->y_charsize * rowdst, vid_priv->xsize, + priv->y_charsize * count); + return 0; }
@@ -348,6 +354,9 @@ static int console_truetype_putc_xy(struct udevice *dev, uint x, uint y,
line += vid_priv->line_length; } + + video_damage(dev->parent, x, y, width, height); + ret = vidconsole_sync_copy(dev, start, line); if (ret) return ret; @@ -415,6 +424,9 @@ static int console_truetype_erase(struct udevice *dev, int xstart, int ystart, } line += vid_priv->line_length; } + + video_damage(dev->parent, xstart, ystart, xend - xstart, yend - ystart); + ret = vidconsole_sync_copy(dev, start, line); if (ret) return ret;
Let's report the video damage when we draw a bitmap on the screen. This way we can later lazily flush only relevant regions to hardware.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer --- drivers/video/video_bmp.c | 2 ++ 1 file changed, 2 insertions(+)
diff --git a/drivers/video/video_bmp.c b/drivers/video/video_bmp.c index 4d2d961696..da8a7b3701 100644 --- a/drivers/video/video_bmp.c +++ b/drivers/video/video_bmp.c @@ -416,6 +416,8 @@ int video_bmp_display(struct udevice *dev, ulong bmp_image, int x, int y, break; };
+ video_damage(dev, x, y, width, height); + /* Find the position of the top left of the image in the framebuffer */ fb = (uchar *)(priv->fb + y * priv->line_length + x * bpix / 8); ret = video_sync_copy(dev, start, fb);
Now that we have a damage tracking API, let's populate damage done by UEFI payloads when they BLT data onto the screen.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer --- lib/efi_loader/efi_gop.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/lib/efi_loader/efi_gop.c b/lib/efi_loader/efi_gop.c index 2c81859807..67286c9a60 100644 --- a/lib/efi_loader/efi_gop.c +++ b/lib/efi_loader/efi_gop.c @@ -33,6 +33,9 @@ struct efi_gop_obj { struct efi_gop ops; struct efi_gop_mode_info info; struct efi_gop_mode mode; +#ifdef CONFIG_DM_VIDEO + struct udevice *vdev; +#endif /* Fields we only have access to during init */ u32 bpix; void *fb; @@ -244,6 +247,10 @@ static __always_inline efi_status_t gop_blt_int(struct efi_gop *this, dlineoff += dwidth; }
+#ifdef CONFIG_DM_VIDEO + video_damage(gopobj->vdev, dx, dy, width, height); +#endif + return EFI_SUCCESS; }
@@ -583,5 +590,9 @@ efi_status_t efi_gop_register(void) gopobj->bpix = bpix; gopobj->fb = fb;
+#ifdef CONFIG_DM_VIDEO + gopobj->vdev = vdev; +#endif + return EFI_SUCCESS; }
On 6/7/22 01:43, Alexander Graf wrote:
Now that we have a damage tracking API, let's populate damage done by UEFI payloads when they BLT data onto the screen.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer
lib/efi_loader/efi_gop.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/lib/efi_loader/efi_gop.c b/lib/efi_loader/efi_gop.c index 2c81859807..67286c9a60 100644 --- a/lib/efi_loader/efi_gop.c +++ b/lib/efi_loader/efi_gop.c @@ -33,6 +33,9 @@ struct efi_gop_obj { struct efi_gop ops; struct efi_gop_mode_info info; struct efi_gop_mode mode; +#ifdef CONFIG_DM_VIDEO
Please, heed the warnings provided by scripts/checkpatch.pl:
WARNING: Use 'if (IS_ENABLED(CONFIG...))' instead of '#if or #ifdef' where possible #174: FILE: lib/efi_loader/efi_gop.c:36: +#ifdef CONFIG_DM_VIDEO
- struct udevice *vdev;
+#endif /* Fields we only have access to during init */ u32 bpix; void *fb; @@ -244,6 +247,10 @@ static __always_inline efi_status_t gop_blt_int(struct efi_gop *this, dlineoff += dwidth; }
+#ifdef CONFIG_DM_VIDEO
WARNING: Use 'if (IS_ENABLED(CONFIG...))' instead of '#if or #ifdef' where possible #184: FILE: lib/efi_loader/efi_gop.c:250: +#ifdef CONFIG_DM_VIDEO
- video_damage(gopobj->vdev, dx, dy, width, height);
+#endif
- return EFI_SUCCESS; }
@@ -583,5 +590,9 @@ efi_status_t efi_gop_register(void) gopobj->bpix = bpix; gopobj->fb = fb;
+#ifdef CONFIG_DM_VIDEO
WARNING: Use 'if (IS_ENABLED(CONFIG...))' instead of '#if or #ifdef' where possible #195: FILE: lib/efi_loader/efi_gop.c:593: +#ifdef CONFIG_DM_VIDEO
Best regards
Heinrich
- gopobj->vdev = vdev;
+#endif
- return EFI_SUCCESS; }
Hey Heinrich,
On 07.06.22 09:12, Heinrich Schuchardt wrote:
On 6/7/22 01:43, Alexander Graf wrote:
Now that we have a damage tracking API, let's populate damage done by UEFI payloads when they BLT data onto the screen.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer
lib/efi_loader/efi_gop.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
diff --git a/lib/efi_loader/efi_gop.c b/lib/efi_loader/efi_gop.c index 2c81859807..67286c9a60 100644 --- a/lib/efi_loader/efi_gop.c +++ b/lib/efi_loader/efi_gop.c @@ -33,6 +33,9 @@ struct efi_gop_obj { struct efi_gop ops; struct efi_gop_mode_info info; struct efi_gop_mode mode; +#ifdef CONFIG_DM_VIDEO
Please, heed the warnings provided by scripts/checkpatch.pl:
WARNING: Use 'if (IS_ENABLED(CONFIG...))' instead of '#if or #ifdef' where possible #174: FILE: lib/efi_loader/efi_gop.c:36: +#ifdef CONFIG_DM_VIDEO
I was mostly afraid of adding a dependency on struct udevice here. But since we already include video.h, I believe we're good. Happy to change it to only runtime checks.
Alex
Now that we have a damage area tells us which parts of the frame buffer actually need updating, let's only dcache flush those on video_sync() calls. With this optimization in place, frame buffer updates - especially on large screen such as 4k displays - speed up significantly.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer --- drivers/video/video-uclass.c | 49 ++++++++++++++++++++++++++++++------ 1 file changed, 42 insertions(+), 7 deletions(-)
diff --git a/drivers/video/video-uclass.c b/drivers/video/video-uclass.c index 9ac1974670..5661beea38 100644 --- a/drivers/video/video-uclass.c +++ b/drivers/video/video-uclass.c @@ -222,6 +222,39 @@ int video_damage(struct udevice *vid, int x, int y, int width, int height) } #endif
+#if defined(CONFIG_ARM) && !CONFIG_IS_ENABLED(SYS_DCACHE_OFF) +static void video_flush_dcache(struct udevice *vid) +{ + struct video_priv *priv = dev_get_uclass_priv(vid); + + if (!priv->flush_dcache) + return; + +#ifdef CONFIG_VIDEO_DAMAGE + if (priv->damage.endx && priv->damage.endy) { + int lstart = priv->damage.x * VNBYTES(priv->bpix); + int lend = priv->damage.endx * VNBYTES(priv->bpix); + int y; + + for (y = priv->damage.y; y < priv->damage.endy; y++) { + ulong fb = (ulong)priv->fb; + ulong start = fb + (y * priv->line_length) + lstart; + ulong end = start + lend; + + start = ALIGN_DOWN(start, CONFIG_SYS_CACHELINE_SIZE); + end = ALIGN(end, CONFIG_SYS_CACHELINE_SIZE); + + flush_dcache_range(start, end); + } + } +#else + flush_dcache_range((ulong)priv->fb, + ALIGN((ulong)priv->fb + priv->fb_size, + CONFIG_SYS_CACHELINE_SIZE)); +#endif +} +#endif + /* Flush video activity to the caches */ int video_sync(struct udevice *vid, bool force) { @@ -240,13 +273,7 @@ int video_sync(struct udevice *vid, bool force) * out whether it exists? For now, ARM is safe. */ #if defined(CONFIG_ARM) && !CONFIG_IS_ENABLED(SYS_DCACHE_OFF) - struct video_priv *priv = dev_get_uclass_priv(vid); - - if (priv->flush_dcache) { - flush_dcache_range((ulong)priv->fb, - ALIGN((ulong)priv->fb + priv->fb_size, - CONFIG_SYS_CACHELINE_SIZE)); - } + video_flush_dcache(vid); #elif defined(CONFIG_VIDEO_SANDBOX_SDL) struct video_priv *priv = dev_get_uclass_priv(vid); static ulong last_sync; @@ -256,6 +283,14 @@ int video_sync(struct udevice *vid, bool force) last_sync = get_timer(0); } #endif + +#ifdef CONFIG_VIDEO_DAMAGE + struct video_priv *priv = dev_get_uclass_priv(vid); + + priv->damage.endx = 0; + priv->damage.endy = 0; +#endif + return 0; }
On 6/7/22 01:43, Alexander Graf wrote:
Now that we have a damage area tells us which parts of the frame buffer actually need updating, let's only dcache flush those on video_sync() calls. With this optimization in place, frame buffer updates - especially on large screen such as 4k displays - speed up significantly.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer
drivers/video/video-uclass.c | 49 ++++++++++++++++++++++++++++++------ 1 file changed, 42 insertions(+), 7 deletions(-)
diff --git a/drivers/video/video-uclass.c b/drivers/video/video-uclass.c index 9ac1974670..5661beea38 100644 --- a/drivers/video/video-uclass.c +++ b/drivers/video/video-uclass.c @@ -222,6 +222,39 @@ int video_damage(struct udevice *vid, int x, int y, int width, int height) } #endif
+#if defined(CONFIG_ARM) && !CONFIG_IS_ENABLED(SYS_DCACHE_OFF)
Why should this be ARM specific?
Best regards
Heinrich
+static void video_flush_dcache(struct udevice *vid) +{
- struct video_priv *priv = dev_get_uclass_priv(vid);
- if (!priv->flush_dcache)
return;+#ifdef CONFIG_VIDEO_DAMAGE
- if (priv->damage.endx && priv->damage.endy) {
int lstart = priv->damage.x * VNBYTES(priv->bpix);int lend = priv->damage.endx * VNBYTES(priv->bpix);int y;for (y = priv->damage.y; y < priv->damage.endy; y++) {ulong fb = (ulong)priv->fb;ulong start = fb + (y * priv->line_length) + lstart;ulong end = start + lend;start = ALIGN_DOWN(start, CONFIG_SYS_CACHELINE_SIZE);end = ALIGN(end, CONFIG_SYS_CACHELINE_SIZE);flush_dcache_range(start, end);}- }
+#else
- flush_dcache_range((ulong)priv->fb,
ALIGN((ulong)priv->fb + priv->fb_size,CONFIG_SYS_CACHELINE_SIZE));+#endif +} +#endif
- /* Flush video activity to the caches */ int video_sync(struct udevice *vid, bool force) {
@@ -240,13 +273,7 @@ int video_sync(struct udevice *vid, bool force) * out whether it exists? For now, ARM is safe. */ #if defined(CONFIG_ARM) && !CONFIG_IS_ENABLED(SYS_DCACHE_OFF)
- struct video_priv *priv = dev_get_uclass_priv(vid);
- if (priv->flush_dcache) {
flush_dcache_range((ulong)priv->fb,ALIGN((ulong)priv->fb + priv->fb_size,CONFIG_SYS_CACHELINE_SIZE));- }
- video_flush_dcache(vid); #elif defined(CONFIG_VIDEO_SANDBOX_SDL) struct video_priv *priv = dev_get_uclass_priv(vid); static ulong last_sync;
@@ -256,6 +283,14 @@ int video_sync(struct udevice *vid, bool force) last_sync = get_timer(0); } #endif
+#ifdef CONFIG_VIDEO_DAMAGE
- struct video_priv *priv = dev_get_uclass_priv(vid);
- priv->damage.endx = 0;
- priv->damage.endy = 0;
+#endif
- return 0; }
On 07.06.22 10:00, Heinrich Schuchardt wrote:
On 6/7/22 01:43, Alexander Graf wrote:
Now that we have a damage area tells us which parts of the frame buffer actually need updating, let's only dcache flush those on video_sync() calls. With this optimization in place, frame buffer updates - especially on large screen such as 4k displays - speed up significantly.
Signed-off-by: Alexander Graf agraf@csgraf.de Reported-by: Da Xue da@libre.computer
drivers/video/video-uclass.c | 49 ++++++++++++++++++++++++++++++------ 1 file changed, 42 insertions(+), 7 deletions(-)
diff --git a/drivers/video/video-uclass.c b/drivers/video/video-uclass.c index 9ac1974670..5661beea38 100644 --- a/drivers/video/video-uclass.c +++ b/drivers/video/video-uclass.c @@ -222,6 +222,39 @@ int video_damage(struct udevice *vid, int x, int y, int width, int height) } #endif +#if defined(CONFIG_ARM) && !CONFIG_IS_ENABLED(SYS_DCACHE_OFF)
Why should this be ARM specific?
I don't believe it should - and that's what the existing comment also says. But currently it is because the dcache API isn't available on all platforms; I'm merely preserving the existing logic :).
Thanks,
Alex
On 6/7/22 01:43, Alexander Graf wrote:
This patch set speeds up graphics output on ARM by a factor of 60x.
On most ARM SBCs, we keep the frame buffer in DRAM and map it as cached, but need it accessible by the display controller which reads directly from a later point of consistency. Hence, we flush the frame buffer to DRAM on every change. The full frame buffer.
Isn't a similar problem already solved by CONFIG_VIDEO_COPY?
Leaving the frame buffer uncached would convert the ARM problem into the X86 case?
Unfortunately, with the advent of 4k displays, we are seeing frame buffers that can take a while to flush out. This was reported by Da Xue with grub, which happily print 1000s of spaces on the screen to draw a menu. Every printed space triggers a cache flush.
This patch set implements the easiest mitigation against this problem: Damage tracking. We remember the lowest common denominator region that was touched since the last video_sync() call and only flush that.
If by "lowest common denominator region" you should mean a rectangle, drawing a point in the upper left corner and another in the lower right corner would require a full flush. So nothing gained in this case.
With this patch set applied, we reduce drawing a large grub menu (with serial console attached for size information) on an RK3399-ROC system at 1440p from 55 seconds to less than 1 second.
Alternatives considered:
Lazy sync - Sandbox does this. It only calls video_sync(true) ever so often. We are missing timers to do this generically.
Double buffering - We could try to identify whether anything changed at all and only draw to the FB if it did. That would require maintaining a second buffer that we need to scan.
Text buffer - Maintain a buffer of all text printed on the screen with respective location. Don't write if the old and new character are identical. This would limit applicability to text only and is an optimization on top of this patch set.
Hash screen lines - Create a hash (sha256?) over every line when it changes. Only flush when it does. I'm not sure if this would waste more time, memory and cache than the current approach. It would make full screen updates much more expensive.
Alexander Graf (6): dm: video: Add damage tracking API dm: video: Add damage notification on display clear vidconsole: Add damage notifications to all vidconsole drivers video: Add damage notification on bmp display efi_loader: GOP: Add damage notification on BLT video: Only dcache flush damaged lines
We need documentation describing the difference between CONFIG_VIDEO_COPY and CONFIG_VIDEO_DAMAGE.
Best regards
Heinrich
drivers/video/Kconfig | 15 ++++++ drivers/video/console_normal.c | 10 ++++ drivers/video/console_rotate.c | 18 +++++++ drivers/video/console_truetype.c | 12 +++++ drivers/video/video-uclass.c | 89 +++++++++++++++++++++++++++++--- drivers/video/video_bmp.c | 2 + include/video.h | 39 +++++++++++++- lib/efi_loader/efi_gop.c | 11 ++++ 8 files changed, 187 insertions(+), 9 deletions(-)
On 07.06.22 10:28, Heinrich Schuchardt wrote:
On 6/7/22 01:43, Alexander Graf wrote:
This patch set speeds up graphics output on ARM by a factor of 60x.
On most ARM SBCs, we keep the frame buffer in DRAM and map it as cached, but need it accessible by the display controller which reads directly from a later point of consistency. Hence, we flush the frame buffer to DRAM on every change. The full frame buffer.
Isn't a similar problem already solved by CONFIG_VIDEO_COPY?
Leaving the frame buffer uncached would convert the ARM problem into the X86 case?
It solves a similar problem, yes. However, it requires us to allocate the frame buffer size twice, and we would need to dynamically toggle the MMU mappings of the frame buffer to WC instead of cached. That's code we don't have today.
VIDEO_COPY is also terribly inefficient in the most common case: Drawing one or multiple characters. It basically copies every line that contains the character, for every character printed. The damage code in this patch set only flushes the relevant rectangles after a string is fully printed.
I think overall, damage tracking with cached memory is simple enough that it gives us the best of all worlds.
Unfortunately, with the advent of 4k displays, we are seeing frame buffers that can take a while to flush out. This was reported by Da Xue with grub, which happily print 1000s of spaces on the screen to draw a menu. Every printed space triggers a cache flush.
This patch set implements the easiest mitigation against this problem: Damage tracking. We remember the lowest common denominator region that was touched since the last video_sync() call and only flush that.
If by "lowest common denominator region" you should mean a rectangle, drawing a point in the upper left corner and another in the lower right corner would require a full flush. So nothing gained in this case.
Glad you asked! :)
While theoretically possible, this is a case that just never happens in U-Boot's code flow. All code that draws to the screen is either blt based (like gop, character drawing or logo display) or moves large portions of the screen (scrolling). The largest granularity we have between syncs is when printing strings. So the worst case you'll have today is a wrap around where you'd end up flushing full lines.
With this patch set applied, we reduce drawing a large grub menu (with serial console attached for size information) on an RK3399-ROC system at 1440p from 55 seconds to less than 1 second.
Alternatives considered:
1) Lazy sync - Sandbox does this. It only calls video_sync(true) ever so often. We are missing timers to do this generically.
2) Double buffering - We could try to identify whether anything changed at all and only draw to the FB if it did. That would require maintaining a second buffer that we need to scan.
3) Text buffer - Maintain a buffer of all text printed on the screen with respective location. Don't write if the old and new character are identical. This would limit applicability to text only and is an optimization on top of this patch set.
4) Hash screen lines - Create a hash (sha256?) over every line when it changes. Only flush when it does. I'm not sure if this would waste more time, memory and cache than the current approach. It would make full screen updates much more expensive.
Alexander Graf (6): dm: video: Add damage tracking API dm: video: Add damage notification on display clear vidconsole: Add damage notifications to all vidconsole drivers video: Add damage notification on bmp display efi_loader: GOP: Add damage notification on BLT video: Only dcache flush damaged lines
We need documentation describing the difference between CONFIG_VIDEO_COPY and CONFIG_VIDEO_DAMAGE.
Hm, maybe we should implement CONFIG_VIDEO_COPY as a flush mechanism behind CONFIG_VIDEO_DAMAGE? That way we only have a single code path for producers left and in addition also optimize drawing individual characters. It would also make the feature useful beyond ARM dcache flushing.
Alex
Date: Thu, 9 Jun 2022 21:04:37 +0200 From: Alexander Graf agraf@csgraf.de
On 07.06.22 10:28, Heinrich Schuchardt wrote:
On 6/7/22 01:43, Alexander Graf wrote:
This patch set speeds up graphics output on ARM by a factor of 60x.
On most ARM SBCs, we keep the frame buffer in DRAM and map it as cached, but need it accessible by the display controller which reads directly from a later point of consistency. Hence, we flush the frame buffer to DRAM on every change. The full frame buffer.
Isn't a similar problem already solved by CONFIG_VIDEO_COPY?
Leaving the frame buffer uncached would convert the ARM problem into the X86 case?
It solves a similar problem, yes. However, it requires us to allocate the frame buffer size twice, and we would need to dynamically toggle the MMU mappings of the frame buffer to WC instead of cached. That's code we don't have today.
For the Apple M1 the framebuffer is covered by the "memory map" and maps it as Normal-NC, but that is because the framebuffer is already set up at the point where u-boot takes control.
Am 9. Juni 2022 21:04:37 MESZ schrieb Alexander Graf agraf@csgraf.de:
On 07.06.22 10:28, Heinrich Schuchardt wrote:
On 6/7/22 01:43, Alexander Graf wrote:
This patch set speeds up graphics output on ARM by a factor of 60x.
On most ARM SBCs, we keep the frame buffer in DRAM and map it as cached, but need it accessible by the display controller which reads directly from a later point of consistency. Hence, we flush the frame buffer to DRAM on every change. The full frame buffer.
Isn't a similar problem already solved by CONFIG_VIDEO_COPY?
Leaving the frame buffer uncached would convert the ARM problem into the X86 case?
It solves a similar problem, yes. However, it requires us to allocate the frame buffer size twice, and we would need to dynamically toggle the MMU mappings of the frame buffer to WC instead of cached. That's code we don't have today.
VIDEO_COPY is also terribly inefficient in the most common case: Drawing one or multiple characters. It basically copies every line that contains the character, for every character printed. The damage code in this patch set only flushes the relevant rectangles after a string is fully printed.
I think overall, damage tracking with cached memory is simple enough that it gives us the best of all worlds.
Unfortunately, with the advent of 4k displays, we are seeing frame buffers that can take a while to flush out. This was reported by Da Xue with grub, which happily print 1000s of spaces on the screen to draw a menu. Every printed space triggers a cache flush.
This patch set implements the easiest mitigation against this problem: Damage tracking. We remember the lowest common denominator region that was touched since the last video_sync() call and only flush that.
If by "lowest common denominator region" you should mean a rectangle, drawing a point in the upper left corner and another in the lower right corner would require a full flush. So nothing gained in this case.
Glad you asked! :)
While theoretically possible, this is a case that just never happens in U-Boot's code flow. All code that draws to the screen is either blt based (like gop, character drawing or logo display) or moves large portions of the screen (scrolling). The largest granularity we have between syncs is when printing strings. So the worst case you'll have today is a wrap around where you'd end up flushing full lines.
With this patch set applied, we reduce drawing a large grub menu (with serial console attached for size information) on an RK3399-ROC system at 1440p from 55 seconds to less than 1 second.
Alternatives considered:
1) Lazy sync - Sandbox does this. It only calls video_sync(true) ever so often. We are missing timers to do this generically.
2) Double buffering - We could try to identify whether anything changed at all and only draw to the FB if it did. That would require maintaining a second buffer that we need to scan.
3) Text buffer - Maintain a buffer of all text printed on the screen with respective location. Don't write if the old and new character are identical. This would limit applicability to text only and is an optimization on top of this patch set.
4) Hash screen lines - Create a hash (sha256?) over every line when it changes. Only flush when it does. I'm not sure if this would waste more time, memory and cache than the current approach. It would make full screen updates much more expensive.
Alexander Graf (6): dm: video: Add damage tracking API dm: video: Add damage notification on display clear vidconsole: Add damage notifications to all vidconsole drivers video: Add damage notification on bmp display efi_loader: GOP: Add damage notification on BLT video: Only dcache flush damaged lines
We need documentation describing the difference between CONFIG_VIDEO_COPY and CONFIG_VIDEO_DAMAGE.
Hm, maybe we should implement CONFIG_VIDEO_COPY as a flush mechanism behind CONFIG_VIDEO_DAMAGE? That way we only have a single code path for producers left and in addition also optimize drawing individual characters. It would also make the feature useful beyond ARM dcache flushing.
Please, consider that RISC-V has no instruction for flushing the data cache. So CONFIG_VIDEO_DAMAGE is not applicable.
Best regards
Heinrich
Alex
On 09.06.22 22:32, Heinrich Schuchardt wrote:
Am 9. Juni 2022 21:04:37 MESZ schrieb Alexander Graf agraf@csgraf.de:
On 07.06.22 10:28, Heinrich Schuchardt wrote:
On 6/7/22 01:43, Alexander Graf wrote:
This patch set speeds up graphics output on ARM by a factor of 60x.
On most ARM SBCs, we keep the frame buffer in DRAM and map it as cached, but need it accessible by the display controller which reads directly from a later point of consistency. Hence, we flush the frame buffer to DRAM on every change. The full frame buffer.
Isn't a similar problem already solved by CONFIG_VIDEO_COPY?
Leaving the frame buffer uncached would convert the ARM problem into the X86 case?
It solves a similar problem, yes. However, it requires us to allocate the frame buffer size twice, and we would need to dynamically toggle the MMU mappings of the frame buffer to WC instead of cached. That's code we don't have today.
VIDEO_COPY is also terribly inefficient in the most common case: Drawing one or multiple characters. It basically copies every line that contains the character, for every character printed. The damage code in this patch set only flushes the relevant rectangles after a string is fully printed.
I think overall, damage tracking with cached memory is simple enough that it gives us the best of all worlds.
Unfortunately, with the advent of 4k displays, we are seeing frame buffers that can take a while to flush out. This was reported by Da Xue with grub, which happily print 1000s of spaces on the screen to draw a menu. Every printed space triggers a cache flush.
This patch set implements the easiest mitigation against this problem: Damage tracking. We remember the lowest common denominator region that was touched since the last video_sync() call and only flush that.
If by "lowest common denominator region" you should mean a rectangle, drawing a point in the upper left corner and another in the lower right corner would require a full flush. So nothing gained in this case.
Glad you asked! :)
While theoretically possible, this is a case that just never happens in U-Boot's code flow. All code that draws to the screen is either blt based (like gop, character drawing or logo display) or moves large portions of the screen (scrolling). The largest granularity we have between syncs is when printing strings. So the worst case you'll have today is a wrap around where you'd end up flushing full lines.
With this patch set applied, we reduce drawing a large grub menu (with serial console attached for size information) on an RK3399-ROC system at 1440p from 55 seconds to less than 1 second.
Alternatives considered:
1) Lazy sync - Sandbox does this. It only calls video_sync(true) ever so often. We are missing timers to do this generically.
2) Double buffering - We could try to identify whether anything changed at all and only draw to the FB if it did. That would require maintaining a second buffer that we need to scan.
3) Text buffer - Maintain a buffer of all text printed on the screen with respective location. Don't write if the old and new character are identical. This would limit applicability to text only and is an optimization on top of this patch set.
4) Hash screen lines - Create a hash (sha256?) over every line when it changes. Only flush when it does. I'm not sure if this would waste more time, memory and cache than the current approach. It would make full screen updates much more expensive.
Alexander Graf (6): dm: video: Add damage tracking API dm: video: Add damage notification on display clear vidconsole: Add damage notifications to all vidconsole drivers video: Add damage notification on bmp display efi_loader: GOP: Add damage notification on BLT video: Only dcache flush damaged lines
We need documentation describing the difference between CONFIG_VIDEO_COPY and CONFIG_VIDEO_DAMAGE.
Hm, maybe we should implement CONFIG_VIDEO_COPY as a flush mechanism behind CONFIG_VIDEO_DAMAGE? That way we only have a single code path for producers left and in addition also optimize drawing individual characters. It would also make the feature useful beyond ARM dcache flushing.
Please, consider that RISC-V has no instruction for flushing the data cache. So CONFIG_VIDEO_DAMAGE is not applicable.
I think we'll have to see what SoCs people come up with. My hope would be that anything that shares DRAM between the display IP block and the CPU speaks on a fully cache coherent bus. Then you don't need any of this trickery.
Alex
participants (4)
-
Alexander Graf -
Heinrich Schuchardt -
Heinrich Schuchardt -
Mark Kettenis