During padding calculation, ox and oy should be ignored unless the
blend operation is neither repeating nor stretching. Otherwise,
the buffer will grow without necessity.
This will allow forcing a specific value for the filter padding,
instead of relying on auto calculation.
Two advantages:
- Auto calculation can't be perfect, since it will add as much
padding as required for the full blur effect
- This prepares the path for animations with effects, where the
object size does not change over time
If the buffer size is smaller than the blurring kernel, then
special precautions must be taken to properly read the source
pixels. Also, fix the corner cases near the left & right edges
(or top & bottom).
When blurring an RGBA buffer to the output buffer,
we don't need to convert the colorspace... but then we'll just
override what was already there.
Introduce a 'dirty' flag set to true whenever a command writes
to an output buffer.
Use two optimizable functions for BOX blur: vertical and horizontal.
These functions will run as many times as requested (from 1 to 6 max).
The horizontal case is pretty straightforward as the source is already
contiguous (nice in terms of cache hits). The only catch is to swap
src and dst without ever writing to the input buffer.
In case of vertical blur, we apply the same method as above, after
rotating the column into a horizontal (contiguous) span, and rotating
it back afterwards.
Now, the same needs to be done for RGBA :)
BOX blur is a lot faster (and easier to optimize, too)
than GAUSSIAN blur. Repeating 2x or 3x BOX blur will also
give similar results to GAUSSIAN blur (very smooth), but
in much less time.
Add a count parameter to the BOX blur instruction.
Actually, there is a very nice trick with BOX blur.
Pass BOX blur 3 times and you can approximate a GAUSSIAN
blur with up to 3% accuracy. This is way more than enough
for just a simple graphical effect.
So, despite the crappy quality of BOX blur, we should
optimize it a lot so we can replace large GAUSSIAN blurs
with series of BOX blurs instead.
Source: Wikipedia's page on box blur :)
This commit also moves around some duplicated definitions.
When a blur operation requires a copy-back to the source
buffer, then the render_op must be set to COPY instead of
BLEND. Otherwise the non blurred content will be visible.
@fix
Prepare optimization paths for blur operations, as they are VERY
costly. This simple change, when using gcc -O3 flag, boosts
horizontal blur performance by > 50%, because STEP is 1 (and
so, memory accesses, increments, etc... are all very simple)
The objective is to have support for NEON, MMX, SSE, too, with
runtime detection.
Remove true Gaussian kernel code, as it is not usable over 12px and
was disabled because it gives different visual results than the
fake Gaussian curve using sin().
According to cedric's horrified comment :)
And add a comment in the code. Yes, this IS a temporary solution,
but the GL engineS being what they are (tons of duplicated code),
I think it's still better for now to just make things work.
Directly use the scale functions from scalecache when
running the GL engine.
With this, and glReadPixels support, the GL engine support is now
100% complete. It will be DAMN SLOW, but ALL filters should work
in both OpenGL and Software rendering.
Make use of glReadPixel to access the source's pixel data.
Use all classic CPU functions to blend and use that data.
Save pointer to the GL image and update it with the latest data
during target render.
Use ENFN's surface_lock, read_pixels, unlock.
Also, add some more error checks to make sure the images are valid,
or return an error at runtime.
This will inform the client whether an asynchronous filter properly
rendered or not.
I actually don't know any case where rendering can fail at runtime.
The structure should not be changed, despite the union modification.
I am renaming for consistency with older branches that had a mask
field in RGBA_Image. Also, the mask.data or data8 is really just
a way to avoid casting between DATA8 and DATA32 (and it shows
clearly what kind of data you are dealing with).
It was possible to keep negative values for dx,dy which would
then draw pixels out of bounds (= crash).
Make check crashed after the previous commit.
@fix
If the filters fail to render at runtime (that is, parsing went fine
but a command failed to run properly), fallback to normal rendering.
This should prevent text from disappearing when using proxies and
the OpenGL engine (for now).
In some situations, text with filters would be rendered in an invalid
position (somewhere too high).
I am not entirely sure of the reason why the original code with BLEND
doesn't work, but this new version is simpler as GL and SW have more
similar behaviours:
- render text to our 'output' buffer
- draw this buffer as an image onto the set target
Thanks zmike for reporting the issue.
And thanks A LOT for using the filters :D
@fix
Signed-off-by: Jean-Philippe Andre <jp.andre@samsung.com>
If a text object changes regularily, there might be cases where
the object will be rendered as a simple black rectangle for just
one frame.
It seems that the previous output buffer is deleted before being
actually rendered on screen. This patch will delay the deletion
of the previous buffer until the current one has been rendered
to the target surface.
And again, thanks zmike for reporting.
@fix
Signed-off-by: Jean-Philippe Andre <jp.andre@samsung.com>
A CRItical message was always displayed when setting a filter
on a text object, saying that proxy rendering is not supported on GL.
Reduce CRI to ERR and skip proxy rendering altogether if there are
no proxy sources.
This @fix needs to be backported.
Thanks zmike for reporting this.
Signed-off-by: Jean-Philippe Andre <jp.andre@samsung.com>
Allow repeat fillmode in blend() for:
alpha --> alpha
alpha --> rgba
rgba --> alpha
Alpha scaling is not implemented yet, but it is not actually
required. Indeed, only proxies can have a different size and
proxies are RGBA images, not alpha.
Alpha scaling may or may not become a requirement in the future,
or for other purposes, but not yet.
When using stretch, all buffers were actually drawn 4 times
on top of each other. This was not visible because in most cases
these buffers were all opaque (alpha = 255 everywhere).
The documentation said color was used as a multiplier, but in
reality the image drawing functions don't use the context's
color when drawing. So the color is only defined for Alpha -> RGBA
operations.
The Windows build (mingw) does not know about strtok_r.
So, let's use the non-safe variant strtok instead.
Currently, this function is called from the main thread only,
so this should be fine :)
In the future it would be nice to not use strtok anymore,
but strtok_r everywhere, and add it to evil. Considering the
release coming soon, I'm not going to change something like that
now.
Test case was:
buffer : a (alpha);
blur (20, dst = a);
blend (src = a, ox = 30);
In that case, padding was 20, 30, 20, 20.
So the blurred buffer was clipped on screen.
If the displacement map has some alpha values (not 0xFF),
then the blending should take this alpha into account. This
part is fine.
BUT, since Evas relies on premultiplied colors... we have a
problem: R (dx) and G (dy) have already been scaled down.
Actually we would need to load the map in non premultiplied RGBA,
otherwise we'll lose precision on dx,dy as soon as A != 0xFF.
Well... I guess this will be a limitation of this filter, for now
at least. Most displacement maps shouldn't even have any alpha
anyways.
In Doxygen format, write the reference documentation for the filters.
It will contain a few examples only, should serve more as a reference
just like edcref.
This is for the script language itself, not for the Eo APIs or the
internal APIs (those are already documented).
Since the transform operation is (for now) a very simple tool,
it only works when src and dst have the same colorspace.
This commit forces users to specify dst, since "input" and "output"
have different colorspaces.
Also, remove globals A, R, G, B from parser.c... these are
temp variables used in a macro.
My CFLAGS didn't include -Wshadow so I missed those.
Thanks Tom for spotting :)
Some features are not supported (mainly because alpha scaling
has not been implemented yet) in the blend API. So, instead of
rendering the wrong effect, fail with an error message.
Also rename col into color for code clarity (we have cols for columns).
Well, proxy sources are rendered to a... GL texture! But we
actually want the image pixels. So we'll need to call glReadPixels
to get them.
Yes, it will be horribly slow. But there isn't really a way around.
This will require a new internal API. For now, just disable the
feature. Hopefully I can make it work soon enough for the release?
If source_set was called after program_set, then parsing would fail.
It used to work because the program was re-parsed at source_set.
Now, save the code, mark the filter as changed, and reparse again
if the source changed (keep track of invalid programs to avoid
excessive parsing).
In async mode, the filter runs in the render thread, so can't
allocate buffers on the fly.
This case should not happen, unless maybe a source has a null
size (eg. it's invisible and not properly rendered).
Proxy sources & objects were not properly unset.
This results either in crashes (especially in the Edje tests)
or dangling objects with tons of references.
Remove the refcount increase/decrease, as it is redundant.
Store pairs proxy+source instead of just the source in all hashes,
so we can unset the is_proxy flag on the proxy when there are no
sources anymore.
Remove compilation warnings: we don't really need cubic
interpolation at this point, we can still add it back
later if wanted.
Also, make it clear that buffer #2 is the output buffer.
Remove meaningless FIXME.
Use the mapped rendering to implement repeat and stretch
with rgba to alpha buffers blending.
If stretch is required, it will add one more (expensive)
scaling step.
This patch implements the final draw from RGBA_Image to the
OpenGL surface. We can even steal the output buffer and
redraw it quickly, without having to re-render everything
(same as in SW).
Since the filters will have to decide on which engine (SW, GL) to
choose from to render the font and the effects, move the font
draw call inside the filters module.
Quick and dirty solution to support the OpenGL engine:
[1] Allocate CPU buffers
[2] Render text and process all effects to these buffers
[3] Push final image as an OpenGL texture.
This patch implements [1].
It is not possible to logically handle padding and offset at the same
time for a proper mirror effect, unless this is handled directly at the
transformation level.
Also, add support for blend() operation padding computation.
This is the simplest solution I can come up with for "mirror" effects.
Displacement maps are HARD to generate and use properly, since the buffer
size is unknown until runtime.
Even if we align the map to the text itself (using the padding information),
it's still hard to describe properly how to apply the displacement map, and
to generate it... So let's just add a simple flip operation.
The displacement effect is way too complicated. Let's keep it
simple and have only one displacement map format (RG + Alpha).
Here's what's missing now:
- Alpha support, to blend in the input with a variable intensity
- Extra padding (see below)
Also, the intensity VS. map values are not perfectly defined yet.
Problems: How to create a complete mirror effect (map needs to go
over boundaries... add extra padding to the buffers).
This is the first possible optimization: save the rendered
text (since we already have the output buffer anyways), and
reuse it if the text + filter didn't change.
Add parameters l, r, t, b to clip the fill area.
While l=x and t=y, the width and height of the clip are determined
at filter run-time, since we don't know the buffer size before.
Brutal method for now: allocate YET ANOTHER buffer,
render scaled image to it (smooth scaling, oh yeah),
use this as a new mask.
For now, supports:
Alpha Input, RGBA mask, RGBA output, X,Y,XY stretching
Syntax was: buffer(name=bla,alpha=bool);
Changed to: buffer:bla(alpha);
There's a semicolon between buffer and its name because ALL whitespaces
are discarded. This might prove useful sometime in the future, so let's
keep it this way for now :)
Padding was brutally calculated by suming ALL the filters'
individual paddings. Now we try to be a bit smarter and propagate
the padding between buffers in the filter chain.
If the buffer is smaller than the blur kernel, then artifacts appear
and CRASHES happen because we read/write out of the buffer bounds.
Output buffer must be larger than the kernel diameter.
Input buffer's size is used to reduce the blurring effect on the edges.
malloc() and mmap() don't return empty buffers, so blending on top
will present tons of artifacts.
Visible mostly on very small buffers as they are malloc()'ed from
previously used memory segments.
So, the (font) effects will be described by a string. It's
basically a new language (yeah yeah sorry), VERY simple, based
on function calls a la Python, with sequential and named arguments.
This string is intended to be passed directly to an evas text object
and embedded into the evas textblock's markup tags.
This file implements both the basic parsing functions, the
compilation of instructions into a queue of commands, and the glue
code for the rest of the filter infrastructure.
These bump maps implement two light effects at once:
- 3D Shadows
- Specular lights
The specular light is activated by a flag.
Another flag enables compensation for darkening/whitening of
horizontal surfaces.
NOTE: This implementation is VERY SLOW.
It uses double values and divisions all over the place.
It might be possible to optimize by computing two LUTs before
running the algorithm.
Displacement maps are a simple but powerful tool to move pixels
around in a buffer, based on a displacement map (image).
Currently, various modes are implemented:
- X, Y or XY displacement
- Alpha map or RGBA map where R and G only are used
An intensity parameter is given as well.
Some of these might not be useful, we can just strip them off when
the final API is decided.
Color curves are a very simple tool to alter the colors of an image,
on a per-pixel basis. This implementation will simply map each pixel
to a 256 bytes buffer, provided by the application.
NOTE: There are no convenience functions yet for easy curve
generation, but this is the plan (give point A, B, B and interpolate
between them to generate the 256 values array).
Currently supported:
- Box blur for Alpha and RGBA
- Gaussian blur for Alpha and RGBA
Motion blur is not implemented.
These effects are SLOW. Gaussian blur is based on a true gaussian
curve for small radii, or a sine-based approximation.
The true gaussian might need to be removed for consistency, since
it gives slightly different results from the sine one (less blur).
It is not possible to mix Alpha and RGBA surfaces in this filter.
These functions differ from normal blending as they will use
a mask on top of input and output buffers.
So, basically they blend input+mask into a third buffer output.
If output is RGBA then mask or input MUST be RGBA.
evas_filter.c contains the entry points for the evas filters subsystem,
in particular, the filter API, the (yet very basic) buffer management
system, the command queue and main context handling functions.
Right now, the Evas engine is left untouched, so these implementations
will work ONLY in the software engine.
NOTE: This will not compile on its own (thus, not added to Makefile.am),
as it will require the filter implementations to be linked.