optimization Tips (for libavcodec):
===================================
What to optimize:
-----------------
If you plan to do non-x86 architecture specific optimizations (SIMD normally),
then take a look in the x86/ directory, as most important functions are
already optimized for MMX.
If you want to do x86 optimizations then you can either try to finetune the
stuff in the x86 directory or find some other functions in the C source to
optimize, but there aren't many left.
Understanding these overoptimized functions:
--------------------------------------------
As many functions tend to be a bit difficult to understand because
of optimizations, it can be hard to optimize them further, or write
architecture-specific versions. It is recommended to look at older
revisions of the interesting files (web frontends for the various FFmpeg
branches are listed at http://ffmpeg.org/download.html).
Alternatively, look into the other architecture-specific versions in
the x86/, ppc/, alpha/ subdirectories. Even if you don't exactly
comprehend the instructions, it could help understanding the functions
and how they can be optimized.
NOTE: If you still don't understand some function, ask at our mailing list!!!
(http://lists.ffmpeg.org/mailman/listinfo/ffmpeg-devel)
When is an optimization justified?
----------------------------------
Normally, clean and simple optimizations for widely used codecs are
justified even if they only achieve an overall speedup of 0.1%. These
speedups accumulate and can make a big difference after awhile. Also, if
none of the following factors get worse due to an optimization -- speed,
binary code size, source size, source readability -- and at least one
factor improves, then an optimization is always a good idea even if the
overall gain is less than 0.1%. For obscure codecs that are not often
used, the goal is more toward keeping the code clean, small, and
readable instead of making it 1% faster.
WTF is that function good for ....:
-----------------------------------
The primary purpose of this list is to avoid wasting time optimizing functions
which are rarely used.
put(_no_rnd)_pixels{,_x2,_y2,_xy2}
Used in motion compensation (en/decoding).
avg_pixels{,_x2,_y2,_xy2}
Used in motion compensation of B-frames.
These are less important than the put*pixels functions.
avg_no_rnd_pixels*
unused
pix_abs16x16{,_x2,_y2,_xy2}
Used in motion estimation (encoding) with SAD.
pix_abs8x8{,_x2,_y2,_xy2}
Used in motion estimation (encoding) with SAD of MPEG-4 4MV only.
These are less important than the pix_abs16x16* functions.
put_mspel8_mc* / wmv2_mspel8*
Used only in WMV2.
it is not recommended that you waste your time with these, as WMV2
is an ugly and relatively useless codec.
mpeg4_qpel* / *qpel_mc*
Used in MPEG-4 qpel motion compensation (encoding & decoding).
The qpel8 functions are used only for 4mv,
the avg_* functions are used only for B-frames.
Optimizing them should have a significant impact on qpel
encoding & decoding.
qpel{8,16}_mc??_old_c / *pixels{8,16}_l4
Just used to work around a bug in an old libavcodec encoder version.
Don't optimize them.
add_bytes/diff_bytes
For huffyuv only, optimize if you want a faster ffhuffyuv codec.
get_pixels / diff_pixels
Used for encoding, easy.
clear_blocks
easiest to optimize
gmc
Used for MPEG-4 gmc.
Optimizing this should have a significant effect on the gmc decoding
speed.
gmc1
Used for chroma blocks in MPEG-4 gmc with 1 warp point
(there are 4 luma & 2 chroma blocks per macroblock, so
only 1/3 of the gmc blocks use this, the other 2/3
use the normal put_pixel* code, but only if there is
just 1 warp point).
Note: DivX5 gmc always uses just 1 warp point.
pix_sum
Used for encoding.
hadamard8_diff / sse / sad == pix_norm1 / dct_sad / quant_psnr / rd / bit
Specific compare functions used in encoding, it depends upon the
command line switches which of these are used.
Don't waste your time with dct_sad & quant_psnr, they aren't
really useful.
put_pixels_clamped / add_pixels_clamped
Used for en/decoding in the IDCT, easy.
Note, some optimized IDCTs have the add/put clamped code included and
then put_pixels_clamped / add_pixels_clamped will be unused.
idct/fdct
idct (encoding & decoding)
fdct (encoding)
difficult to optimize
dct_quantize_trellis
Used for encoding with trellis quantization.
difficult to optimize
dct_quantize
Used for encoding.
dct_unquantize_mpeg1
Used in MPEG-1 en/decoding.
dct_unquantize_mpeg2
Used in MPEG-2 en/decoding.
dct_unquantize_h263
Used in MPEG-4/H.263 en/decoding.
Alignment:
Some instructions on some architectures have strict alignment restrictions,
for example most SSE/SSE2 instructions on x86.
The minimum guaranteed alignment is written in the .h files, for example:
void (*put_pixels_clamped)(const int16_t *block/*align 16*/, UINT8 *pixels/*align 8*/, int line_size);
General Tips:
-------------
Use asm loops like:
__asm__(
"1: ....
...
"jump_instruction ....
Do not use C loops:
do{
__asm__(
...
}while()
For x86, mark registers that are clobbered in your asm. This means both
general x86 registers (e.g. eax) as well as XMM registers. This last one is
particularly important on Win64, where xmm6-15 are callee-save, and not
restoring their contents leads to undefined results. In external asm (e.g.
yasm), you do this by using:
cglobal functon_name, num_args, num_regs, num_xmm_regs
In inline asm, you specify clobbered registers at the end of your asm:
__asm__(".." ::: "%eax").
If gcc is not set to support sse (-msse) it will not accept xmm registers
in the clobber list. For that we use two macros to declare the clobbers.
XMM_CLOBBERS should be used when there are other clobbers, for example:
__asm__(".." ::: XMM_CLOBBERS("xmm0",) "eax");
and XMM_CLOBBERS_ONLY should be used when the only clobbers are xmm registers:
__asm__(".." :: XMM_CLOBBERS_ONLY("xmm0"));
Do not expect a compiler to maintain values in your registers between separate
(inline) asm code blocks. It is not required to. For example, this is bad:
__asm__("movdqa %0, %%xmm7" : src);
/* do something */
__asm__("movdqa %%xmm7, %1" : dst);
- first of all, you're assuming that the compiler will not use xmm7 in
between the two asm blocks. It probably won't when you test it, but it's
a poor assumption that will break at some point for some --cpu compiler flag
- secondly, you didn't mark xmm7 as clobbered. If you did, the compiler would
have restored the original value of xmm7 after the first asm block, thus
rendering the combination of the two blocks of code invalid
Code that depends on data in registries being untouched, should be written as
a single __asm__() statement. Ideally, a single function contains only one
__asm__() block.
Use external asm (nasm/yasm) or inline asm (__asm__()), do not use intrinsics.
The latter requires a good optimizing compiler which gcc is not.
When debugging a x86 external asm compilation issue, if lost in the macro
expansions, add DBG=1 to your make command-line: the input file will be
preprocessed, stripped of the debug/empty lines, then compiled, showing the
actual lines causing issues.
Inline asm vs. external asm
---------------------------
Both inline asm (__asm__("..") in a .c file, handled by a compiler such as gcc)
and external asm (.s or .asm files, handled by an assembler such as yasm/nasm)
are accepted in FFmpeg. Which one to use differs per specific case.
- if your code is intended to be inlined in a C function, inline asm is always
better, because external asm cannot be inlined
- if your code calls external functions, yasm is always better
- if your code takes huge and complex structs as function arguments (e.g.
MpegEncContext; note that this is not ideal and is discouraged if there
are alternatives), then inline asm is always better, because predicting
member offsets in complex structs is almost impossible. It's safest to let
the compiler take care of that
- in many cases, both can be used and it just depends on the preference of the
person writing the asm. For new asm, the choice is up to you. For existing
asm, you'll likely want to maintain whatever form it is currently in unless
there is a good reason to change it.
- if, for some reason, you believe that a particular chunk of existing external
asm could be improved upon further if written in inline asm (or the other
way around), then please make the move from external asm <-> inline asm a
separate patch before your patches that actually improve the asm.
Links:
======
http://www.aggregate.org/MAGIC/
x86-specific:
-------------
http://developer.intel.com/design/pentium4/manuals/248966.htm
The IA-32 Intel Architecture Software Developer's Manual, Volume 2:
Instruction Set Reference
http://developer.intel.com/design/pentium4/manuals/245471.htm
http://www.agner.org/assem/
AMD Athlon Processor x86 Code Optimization Guide:
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22007.pdf
ARM-specific:
-------------
ARM Architecture Reference Manual (up to ARMv5TE):
http://www.arm.com/community/university/eulaarmarm.html
Procedure Call Standard for the ARM Architecture:
http://www.arm.com/pdfs/aapcs.pdf
Optimization guide for ARM9E (used in Nokia 770 Internet Tablet):
http://infocenter.arm.com/help/topic/com.arm.doc.ddi0240b/DDI0240A.pdf
Optimization guide for ARM11 (used in Nokia N800 Internet Tablet):
http://infocenter.arm.com/help/topic/com.arm.doc.ddi0211j/DDI0211J_arm1136_r1p5_trm.pdf
Optimization guide for Intel XScale (used in Sharp Zaurus PDA):
http://download.intel.com/design/intelxscale/27347302.pdf
Intel Wireless MMX 2 Coprocessor: Programmers Reference Manual
http://download.intel.com/design/intelxscale/31451001.pdf
PowerPC-specific:
-----------------
PowerPC32/AltiVec PIM:
www.freescale.com/files/32bit/doc/ref_manual/ALTIVECPEM.pdf
PowerPC32/AltiVec PEM:
www.freescale.com/files/32bit/doc/ref_manual/ALTIVECPIM.pdf
CELL/SPU:
http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/30B3520C93F437AB87257060006FFE5E/$file/Language_Extensions_for_CBEA_2.4.pdf
http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/9F820A5FFA3ECE8C8725716A0062585F/$file/CBE_Handbook_v1.1_24APR2007_pub.pdf
GCC asm links:
--------------
official doc but quite ugly
http://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
a bit old (note "+" is valid for input-output, even though the next disagrees)
http://www.cs.virginia.edu/~clc5q/gcc-inline-asm.pdf