sw_engine: radial gradient
[issues 2014: radial gradient](#2014)
Radial gradient results in a corrupted image when the focal point is outside the circle on Apple Silicon.
This happens because some compilers use FMA to optimize the a = dr² - dx² - dy² calculation,
which cause loss of precision.
We rely on temporary variables to prevent FMA.
We could also use compiler specific float contraction control pragmas to avoid this if this doesn't work in the future
After introducing the length measurement
of approximate and exact Bezier curves in
a8c0030d80
the calculation of the outline length still
used the approximate functions, while further
calculations took the exact value into account.
This discrepancy led to the creation of artifacts.
@Issue: https://github.com/thorvg/thorvg/issues/2021
stroke line drawing has been broken at a certain case,
this reverts a part of change from the old optimization:
d81f5d29fb
note that this change wouldn't affect any performance.
issue: https://github.com/thorvg/thorvg/issues/2015
Dashed curves require greater precision in calculating
their lengths and while splitting Bezier curves.
Otherwise, it results in visual discrepancies compared
to the expected outcomes.
Approximate functions 'bezLengthApprox' and 'bezAtApprox'
used for calculations in the lottie loader.
issue: https://github.com/thorvg/thorvg/issues/1686
In the case of non-uniform scaling for
thick strokes, artifacts were visible.
The calculations took into account the angle
resulting from the already transformed points,
whereas the untransformed points should have
been considered - the transformation is taken
into account in the next step.
@issue: https://github.com/thorvg/thorvg/issues/1915
Adjusted the sampling position basis by -0.5 pixel
to enhance edge line interpolation quality.
This change addresses an issue with the image up-scaling process,
resulting in clearer and more accurate visuals.
issue: https://github.com/thorvg/thorvg/issues/1949
Previously, the logic was set to halt rendering when any part of the paints,
particularly bitmap-based images, failed to render.
This update modifies the behavior to continue drawing the scene,
allowing for the successful rendering of other elements.
issue: https://github.com/thorvg/thorvg/issues/1951
This update fixes an issue where duplicated overlapping curves/lines,
when introduced as a single line.
That case overlapping command operation is take account for
the trim path's path-building process.
We now skip such cases to ensure the trim path functions correctly.
issue: https://github.com/thorvg/thorvg/issues/1939
resolved a corner case where assets were being attached
multiple times during scene building. This issue was causing
a significant performence drop in image animation
in certain scenarios.
The compositor memory is likely to be reused in the next frame.
To enhance performance, it is advisable to retain this memory by default.
We may consider introducing a cache policy interface in the Initializer.
This would allow users to manage the Canvas memory more effectively.
Anyhow, this improves the Lottie example performance by 10%
Adjust the sampling count according to the scale ratio.
This significantly improves performance
while making it hard to recognize any loss in image quality.
Lottie example performance has improved by 15%.
Some systems such as micro-processor might not support
the thread feature on the system.
Enhance the portability by compiling the thorvg with toggling the
threading depepdency through the build option.
For this, thorvg newly introduced the internal Key/ScopedLock abstraction
for transparent thread-locking dependnecy.
To turn off the thread feature, please use the next build option:
$meson setup build -Dthreads=false ...
Note that, the thread feature is enabled in default.
Turning off the thread feature could reduce the binary size by 7kb.
issue: https://github.com/thorvg/thorvg/issues/1900
Introduced a dedicated mutex for each surface instance
to ensure safe sharing between the loader, renderer, and engine.
This enhancement allows for secure modification and access to bitmap data,
addressing potential concurrency issues.
Multiple Picture instances can now safely share a single loader instance,
optimizing performance.
This change builds upon the previous Loader Cache improvements:
ff6ea4b6c4
Basic shapes were trimmed entirely when they were outside of the canvas,
even if they had a big enough stroke to be partially on the canvas.
This fixes the issue.
Issue: https://github.com/thorvg/thorvg/issues/1785
Unified common logic for scaled image raster operations,
Avoid on-spot pixel computation as possible.
Tested on local machine (single thread)
Lottie: 0.057s -> 0.053s (-0.004s)
Try to minimize the use of sqrt() and arctan() calls
when possible. These calls can be relatively expensive
when accumulated within a single frame.
Also repalce the division with shift operation.
since split cubic function is one of the significant hot-spots
in the data processing, we could earn a noticable enhancement.
Tested on single thread local machine:
Lottie: 0.080 -> 0.052s (-0.028s)
Performance: 0.023 -> 0.022 (-0.001s)
Binary: +34
Streamlining computations with floating-point operations in rotation
thereby improving 'thorvg' speed.
Also use the well-optimized posix math functions instead of
custom math.
Test on my local machine.
Lottie: -0.008s (0.073 -> 0.065)
Performance: -0.0013s (0.0154 -> 0.0141)
Binary: -323
Previously, the engine didn't properly cover the dash line corner styles
because it considered a new line to start at the corner.
This update modifies the logic to recognize curved lines
as a single line, including the corners.
There may still be some quality issues,
but it's an improvement over the previous version.
@Issue: https://github.com/thorvg/thorvg/issues/121
The color conversion is supposed to take into account the differences between
straight alpha premultiplied color and pre-multiplied alpha color.
The previous logic does not perfectly cover these conditions.
The problem was occured in the thorvg viewer with a jpeg bgra format.
While it may represent a dot with a stroke width,
ignoring its length is not a good idea.
this also eliminates 0 size dash dots caused by unexpected empty lengths.
