The basic idea is to break down normal repeat into a set of
non-repeat scanline compositions and stitching them together.
Bilinear may interpolate last and first pixels of source scanline.
In this case, we can use temporary wrap around buffer.
By replacing boolean arguments with flags, the code can be more
readable and flags can be extended to do some more things later.
Currently following flags are defined.
FLAG_NONE
- No flags are turned on.
FLAG_HAVE_SOLID_MASK
- Template will generate solid mask composite functions.
FLAG_HAVE_NON_SOLID_MASK
- Template will generate bits mask composite functions.
FLAG_HAVE_SOLID_MASK and FLAG_NON_SOLID_MASK should be mutually
exclusive.
The first bug is that a vmull.u8 instruction would store its result in
the q1 register, clobbering the d2 register used later on. The second
is that a vraddhn instruction would overwrite d25, corrupting the q12
register used later.
Fixing the second bug caused a pipeline bubble where the d18 register
would be unavailable for a clock cycle. This is fixed by swapping the
instruction with its successor.
Move the eight most common formats to the top of the list of image
formats and make create_random_image() much more likely to select one
of those eight formats.
This should help catch more bugs in SIMD optimized operations.
Autoconf 2.86 reports:
warning: AC_LANG_CONFTEST: no AC_LANG_SOURCE call detected in body
Every code fragment must be wrapped in [AC_LANG_SOURCE([...])]
The variables in question were dst_x, dst_y, dst_image. The majority
of _x and _y uses were already dest_x and dest_y, while the majority
of _image uses were dst_image.
When the image is a8r8g8b8 and not transformed, and the fetched
rectangle is within the image bounds, scanlines can be fetched by
simply returning a pointer instead of copying the bits.
It will at some point become useful to have CPU specific destination
iterators. However, a problem with that, is that such iterators should
not be used if we can composite directly in the destination image.
By moving the noop destination iterator to the noop implementation, we
can ensure that it will be chosen before any CPU specific iterator.
The DST operator doesn't actually do anything, so add a noop "fast
path" for it, instead of checking in pixman_image_composite32().
The performance tradeoff here is that we get rid of a test for DST in
the common case where the operator is not DST, in return for an extra
walk over the clip rectangles in the uncommon case where the operator
actually is DST.
This new implementation is ahead of all other implementations in the
fallback chain and is supposed to contain operations that are "noops",
ie., they don't require any work. For example, it might contain a
"fast path" for the DST operator that doesn't actually do anything or
an iterator for a8r8g8b8 that just returns a pointer into the image.
General fetch->combine->store based bilinear scanline functions.
Need further optimizations and eventually will be replaced with optimal
functions one by one.
General functions should be located in pixman-arm-neon-asm-bilinear.S and
optimal functions in pixman-arm-neon-asm.S
Following general bilinear scanline functions are implemented
over_8888_8888
add_8888_8888
src_8888_8_8888
src_8888_8_0565
src_0565_8_x888
src_0565_8_0565
over_8888_8_8888
add_8888_8_8888
Previously, this function would do coordinate calculations in such a
way that (x_dst, y_dst) would only affect the alignment of the source
image, but not of the traps, which would always be considered to be in
absolute destination coordinates. This is unlike the
pixman_image_composite() function which also registers the mask to the
destination.
This patch makes it so that traps are also offset by (x_dst, y_dst).
Also add a comment explaining how this function is supposed to
operate, and update tri-test.c and composite-trap-test.c to deal with
the new semantics.
This improves the performance of the firefox-talos-gfx benchmark with
the image16 backend. Benchmark on an 800 MHz ARM Cortex A8:
Before:
[ # ] backend test min(s) median(s) stddev. count
[ 0] image16 firefox-talos-gfx 121.773 122.218 0.15% 6/6
After:
[ # ] backend test min(s) median(s) stddev. count
[ 0] image16 firefox-talos-gfx 85.247 85.563 0.22% 6/6
V2: Slightly better instruction scheduling based on comments from Taekyun Kim.
V3: Eliminate all stalls from the inner loop. Also based on comments from Taekyun Kim.
PIXMAN_LINK_WITH_ENV did not fail unless -Wall -Werror is used.
So even when the compiler did not support OpenMP, USE_OPENMP was defined.
Fix that by running the second OpenMP test only when first AC_OPENMP find supported
configure tested in the cases :
gcc without libgomp support, no openmp option, --enable-openmp and --disable-openmp
gcc with libgomp support, no openmp option, --enable-openmp and --disable-openmp
Not tested with autoconf version not knowing openmp (<2.62)
Warn when --enable-openmp is requested but no support is found
Signed-off-by: Gilles Espinasse <g.esp@free.fr>
Performance of the inner loop when working with the data in L1 cache:
ARM Cortex-A8: 41 cycles per 4 pixels (no stalls and partial dual issue)
ARM Cortex-A9: 48 cycles per 4 pixels (no stalls)
It might be still possible to improve performance even more on ARM Cortex-A8
with a better use of dual issue.
Benchmark on ARM Cortex-A8 r1p3 @600MHz, 32-bit LPDDR @166MHz:
Microbenchmark (scaling 2000x2000 image with scale factor close to 1x):
before: op=1, src=20028888, dst=20028888, speed=40.38 MPix/s
after: op=1, src=20028888, dst=20028888, speed=48.47 MPix/s
Benchmark on ARM Cortex-A8 r2p2 @1GHz, 32-bit LPDDR @200MHz:
Microbenchmark (scaling 2000x2000 image with scale factor close to 1x):
before: op=1, src=20028888, dst=20028888, speed=79.68 MPix/s
after: op=1, src=20028888, dst=20028888, speed=93.11 MPix/s
Now an extra 'flag' parameter is supported in bilinear scaline scaling
function generation macro. It can be used to enable 4 or 8 pixels per
loop iteration unrolling and provide save/restore code for d8-d15
registers.
