Always use the special filter for the first and last 3 columns (only).
Changes made in 64ed397 slowed the filter to just under 3/4 of what it
was. This commit restores the speed while maintaining identical output.
For reference, on my Athlon64:
1733222 decicycles in old
2358563 decicycles in new
1727558 decicycles in this
Signed-off-by: Anton Khirnov <anton@khirnov.net>
Current code divides before increasing precision.
Also reduce upper bound for strength from 255 to 64. This will prevent
an overflow in the SSSE3 and MMX filter_line code: delta is expressed as
an u16 being shifted by 2 to the left. If it overflows, having a
strength not above 64 will make sure that m is set to 0 (making the
m*m*delta >> 14 expression void).
A value above 64 should not make any sense unless gradfun is used as
a blur filter.
Signed-off-by: Anton Khirnov <anton@khirnov.net>
These smaller samples do not need to be unpacked to double words
allowing the code to process more pixels every iteration (still 2 in MMX
but 6 in SSE2). It also avoids emulating the missing double word
instructions on older instruction sets.
Like with the previous code for 16-bit samples this has been tested on
an Athlon64 and a Core2Quad.
Athlon64:
1809275 decicycles in C, 32718 runs, 50 skips
911675 decicycles in mmx, 32727 runs, 41 skips, 2.0x faster
495284 decicycles in sse2, 32747 runs, 21 skips, 3.7x faster
Core2Quad:
921363 decicycles in C, 32756 runs, 12 skips
486537 decicycles in mmx, 32764 runs, 4 skips, 1.9x faster
293296 decicycles in sse2, 32759 runs, 9 skips, 3.1x faster
284910 decicycles in ssse3, 32759 runs, 9 skips, 3.2x faster
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
This is a fairly dumb copy of the assembly for 8-bit samples but it
works and produces identical output to the C version. The options have
been tested on an Athlon64 and a Core2Quad.
Athlon64:
1810385 decicycles in C, 32726 runs, 42 skips
1080744 decicycles in mmx, 32744 runs, 24 skips, 1.7x faster
818315 decicycles in sse2, 32735 runs, 33 skips, 2.2x faster
Core2Quad:
924025 decicycles in C, 32750 runs, 18 skips
623995 decicycles in mmx, 32767 runs, 1 skips, 1.5x faster
406223 decicycles in sse2, 32764 runs, 4 skips, 2.3x faster
387842 decicycles in ssse3, 32767 runs, 1 skips, 2.4x faster
307726 decicycles in sse4, 32763 runs, 5 skips, 3.0x faster
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Always use the special filter for the first and last 3 columns (only).
Changes made in 64ed397 slowed the filter to just under 3/4 of what it
was. This commit restores the speed while maintaining identical output.
For reference, on my Athlon64:
1733222 decicycles in old
2358563 decicycles in new
1727558 decicycles in this
Signed-off-by: Michael Niedermayer <michaelni@gmx.at>
Manually load registers to avoid using 8 registers on x86_32 with
compilers that do not align the stack (e.g. MSVC).
Signed-off-by: Diego Biurrun <diego@biurrun.de>
Under some circumstances, suncc will use a single register for the
address of all memory operands, inserting lea instructions loading
the correct address prior to each memory operand being used in the
code. In the yadif code, the branch in the asm block bypasses such
an lea instruction, causing an incorrect address to be used in the
following load.
This patch replaces the tmpX arrays with a single array and uses a
register operand to hold its address. Although this prevents using
offsets from the stack pointer to access these locations, the code
still builds as 32-bit PIC even with old compilers.
Signed-off-by: Mans Rullgard <mans@mansr.com>
Refactoring mmx2/mmxext YASM code with cpuflags will force renames.
So switching to a consistent naming scheme beforehand is sensible.
The name "mmxext" is more official and widespread and also the name
of the CPU flag, as reported e.g. by the Linux kernel.
Paul B Mahol
Diego Biurrun
Clément Bœsch
James Darnley
Loren Merritt
Anton Khirnov
Michael Niedermayer
Daniel Kang
Carl Eugen Hoyos
Justin Ruggles
Martin Storsjö
Mans Rullgard
Ronald S. Bultje