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.
Moving horizontal interpolation weights update instructions from the
beginning of loop to its end allows to hide some pipeline stalls and
improve performance.
Instead of two
mvn d24, d24
mvn d25, d25
use just one
mvn q12, q12
Also move another vmvn instruction into the created pipeline bubble,
as pointed out by Siarhei.
Up until now, all pixman release, both snapshots and releases were
uploaded to the "releases" directory on www.cairographics.org, but
it's better to development snapshots in the "snapshots" directory.
This patch changes Makefile.am to do that.
When run in PIXMAN_RANDOMIZE_TESTS mode, this test would go into an
infinite loop because the loop started at 'seed' but the stop
condition was still N_TESTS.
This format is particularly useful on big-endian architectures, where RGBA in
memory/file order corresponds to r8g8b8a8 as an uint32_t. This is important
because RGBA is in some cases the only available choice (for example as a pixel
format in OpenGL ES 2.0).
This patch makes so that composite and stress-test will start from a
random seed if the PIXMAN_RANDOMIZE_TESTS environment variable is
set. Running the test suite in this mode is useful to get more test
coverage.
Also, in stress-test.c make it so that setting the initial seed causes
threads to be turned off. This makes it much easier to see when
something fails.
All of the information previously passed to the iterator initializers
is now available in the iterator itself, so there is no need to pass
it as arguments anymore.
This makes _pixman_implementation_{src,dest}_iter_init() responsible
for filling parts of the information in the iterators. Specifically,
the information passed as arguments is stored in the iterator.
Also add a height field to pixman_iter_t().
There is no reason to go through
_pixman_implementation_{src,dest}_iter_init(), especially since
_pixman_implementation_src_iter_init() is doing various other checks
that only need to be done once.
Also call delegate->src_iter_init() directly in pixman-sse2.c
Instructions scheduling improved in the code responsible for fetching r5g6b5
pixels and converting them to the intermediate x8r8g8b8 color format used in
the interpolation part of code. Still a lot of NEON stalls are remaining,
which can be resolved later by the use of pipelining.
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=10020565, dst=10020565, speed=32.29 MPix/s
op=1, src=10020565, dst=20020888, speed=36.82 MPix/s
after: op=1, src=10020565, dst=10020565, speed=41.35 MPix/s
op=1, src=10020565, dst=20020888, speed=49.16 MPix/s
This is a cleanup for old and now duplicated code. The performance improvement
is mostly coming from the enabled use of software prefetch, but instructions
scheduling is also slightly better.
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=53.24 MPix/s
after: op=1, src=20028888, dst=20028888, speed=74.36 MPix/s
This allows to generate bilinear scanline scaling functions targeting
various source and destination color formats. Right now a8r8g8b8/x8r8g8b8
and r5g6b5 color formats are supported. More formats can be added if needed.
Initial NEON optimization for bilinear scaling. Can be probably
improved more.
Benchmark on ARM Cortex-A8:
Microbenchmark (scaling 2000x2000 image with scale factor close to 1x):
before: op=1, src=20028888, dst=20028888, speed=6.70 MPix/s
after: op=1, src=20028888, dst=20028888, speed=44.27 MPix/s
A primitive naive implementation of bilinear scaling using SSE2 intrinsics,
which only handles one pixel at a time. It is approximately 2x faster than
pixman general compositing path. Single pass processing without intermediate
temporary buffer contributes to ~15% and loop unrolling contributes to ~20%
of this speedup.
Benchmark on Intel Core i7 (x86-64):
Using cairo-perf-trace:
before: image firefox-planet-gnome 12.566 12.610 0.23% 6/6
after: image firefox-planet-gnome 10.961 11.013 0.19% 5/6
Microbenchmark (scaling 2000x2000 image with scale factor close to 1x):
before: op=1, src=20028888, dst=20028888, speed=70.48 MPix/s
after: op=1, src=20028888, dst=20028888, speed=165.38 MPix/s
Individual correctness check for the new bilinear scaling related
supplementary function. This test program uses a bit wider range
of input arguments, not covered by other tests.
Can be used for implementing SIMD optimized fast path
functions which work with bilinear scaled source images.
Similar to the template for nearest scaling main loop, the
following types of mask are supported:
1. no mask
2. non-scaled a8 mask with SAMPLES_COVER_CLIP flag
3. solid mask
PAD repeat is fully supported. NONE repeat is partially
supported (right now only works if source image has alpha
channel or when alpha channel of the source image does not
have any effect on the compositing operation).
MSVC does not notice non-returning functions (abort() / assert(0))
and warns about paths which end with them in non-void functions:
c:\cygwin\home\ranma42\code\fdo\pixman\test\fetch-test.c(114) :
warning C4715: 'reader' : not all control paths return a value
c:\cygwin\home\ranma42\code\fdo\pixman\test\stress-test.c(133) :
warning C4715: 'real_reader' : not all control paths return a value
c:\cygwin\home\ranma42\code\fdo\pixman\test\composite.c(431) :
warning C4715: 'calc_op' : not all control paths return a value
These warnings can be silenced by adding a return after the
termination call.
