Statistics
| Branch: | Revision:

ffmpeg / doc / optimization.txt @ 8f738eea

History | View | Annotate | Download (6.99 KB)

1
optimization Tips (for libavcodec):
2

    
3
What to optimize:
4
If you plan to do non-x86 architecture specific optimizations (SIMD normally),
5
then take a look in the i386/ directory, as most important functions are
6
already optimized for MMX.
7

    
8
If you want to do x86 optimizations then you can either try to finetune the
9
stuff in the i386 directory or find some other functions in the C source to
10
optimize, but there aren't many left.
11

    
12
Understanding these overoptimized functions:
13
As many functions tend to be a bit difficult to understand because
14
of optimizations, it can be hard to optimize them further, or write
15
architecture-specific versions. It is recommened to look at older
16
revisions of the interesting files (for a web frontend try ViewVC at
17
http://svn.mplayerhq.hu/ffmpeg/trunk/).
18
Alternatively, look into the other architecture-specific versions in
19
the i386/, ppc/, alpha/ subdirectories. Even if you don't exactly
20
comprehend the instructions, it could help understanding the functions
21
and how they can be optimized.
22

    
23
NOTE: If you still don't understand some function, ask at our mailing list!!!
24
(http://lists.mplayerhq.hu/mailman/listinfo/ffmpeg-devel)
25

    
26
What speedup justifies an optimizetion?
27
Normaly with clean&simple optimizations and widely used codecs a overall
28
speedup of the affected codec of 0.1% is enough. These speedups accumulate
29
and can make a big difference after a while ...
30
Also if none of the following gets worse and at least one gets better then an
31
optimization is always a good idea even if the overall gain is less than 0.1%
32
(speed, binary code size, source size, source readability)
33
For obscure codecs noone uses, the goal is more toward keeping the code clean
34
small and readable than to make it 1% faster.
35

    
36

    
37
WTF is that function good for ....:
38
The primary purpose of that list is to avoid wasting time to optimize functions
39
which are rarely used
40

    
41
put(_no_rnd)_pixels{,_x2,_y2,_xy2}
42
    Used in motion compensation (en/decoding).
43

    
44
avg_pixels{,_x2,_y2,_xy2}
45
    Used in motion compensation of B-frames.
46
    These are less important than the put*pixels functions.
47

    
48
avg_no_rnd_pixels*
49
    unused
50

    
51
pix_abs16x16{,_x2,_y2,_xy2}
52
    Used in motion estimation (encoding) with SAD.
53

    
54
pix_abs8x8{,_x2,_y2,_xy2}
55
    Used in motion estimation (encoding) with SAD of MPEG-4 4MV only.
56
    These are less important than the pix_abs16x16* functions.
57

    
58
put_mspel8_mc* / wmv2_mspel8*
59
    Used only in WMV2.
60
    it is not recommended that you waste your time with these, as WMV2
61
    is an ugly and relatively useless codec.
62

    
63
mpeg4_qpel* / *qpel_mc*
64
    Used in MPEG-4 qpel motion compensation (encoding & decoding).
65
    The qpel8 functions are used only for 4mv,
66
    the avg_* functions are used only for B-frames.
67
    Optimizing them should have a significant impact on qpel
68
    encoding & decoding.
69

    
70
qpel{8,16}_mc??_old_c / *pixels{8,16}_l4
71
    Just used to work around a bug in an old libavcodec encoder version.
72
    Don't optimize them.
73

    
74
tpel_mc_func {put,avg}_tpel_pixels_tab
75
    Used only for SVQ3, so only optimize them if you need fast SVQ3 decoding.
76

    
77
add_bytes/diff_bytes
78
    For huffyuv only, optimize if you want a faster ffhuffyuv codec.
79

    
80
get_pixels / diff_pixels
81
    Used for encoding, easy.
82

    
83
clear_blocks
84
    easiest to optimize
85

    
86
gmc
87
    Used for MPEG-4 gmc.
88
    Optimizing this should have a significant effect on the gmc decoding
89
    speed.
90

    
91
gmc1
92
    Used for chroma blocks in MPEG-4 gmc with 1 warp point
93
    (there are 4 luma & 2 chroma blocks per macroblock, so
94
    only 1/3 of the gmc blocks use this, the other 2/3
95
    use the normal put_pixel* code, but only if there is
96
    just 1 warp point).
97
    Note: DivX5 gmc always uses just 1 warp point.
98

    
99
pix_sum
100
    Used for encoding.
101

    
102
hadamard8_diff / sse / sad == pix_norm1 / dct_sad / quant_psnr / rd / bit
103
    Specific compare functions used in encoding, it depends upon the
104
    command line switches which of these are used.
105
    Don't waste your time with dct_sad & quant_psnr, they aren't
106
    really useful.
107

    
108
put_pixels_clamped / add_pixels_clamped
109
    Used for en/decoding in the IDCT, easy.
110
    Note, some optimized IDCTs have the add/put clamped code included and
111
    then put_pixels_clamped / add_pixels_clamped will be unused.
112

    
113
idct/fdct
114
    idct (encoding & decoding)
115
    fdct (encoding)
116
    difficult to optimize
117

    
118
dct_quantize_trellis
119
    Used for encoding with trellis quantization.
120
    difficult to optimize
121

    
122
dct_quantize
123
    Used for encoding.
124

    
125
dct_unquantize_mpeg1
126
    Used in MPEG-1 en/decoding.
127

    
128
dct_unquantize_mpeg2
129
    Used in MPEG-2 en/decoding.
130

    
131
dct_unquantize_h263
132
    Used in MPEG-4/H.263 en/decoding.
133

    
134
FIXME remaining functions?
135
BTW, most of these functions are in dsputil.c/.h, some are in mpegvideo.c/.h.
136

    
137

    
138

    
139
Alignment:
140
Some instructions on some architectures have strict alignment restrictions,
141
for example most SSE/SSE2 instructions on x86.
142
The minimum guaranteed alignment is written in the .h files, for example:
143
    void (*put_pixels_clamped)(const DCTELEM *block/*align 16*/, UINT8 *pixels/*align 8*/, int line_size);
144

    
145

    
146

    
147
Links:
148
http://www.aggregate.org/MAGIC/
149

    
150
x86-specific:
151
http://developer.intel.com/design/pentium4/manuals/248966.htm
152

    
153
The IA-32 Intel Architecture Software Developer's Manual, Volume 2:
154
Instruction Set Reference
155
http://developer.intel.com/design/pentium4/manuals/245471.htm
156

    
157
http://www.agner.org/assem/
158

    
159
AMD Athlon Processor x86 Code Optimization Guide:
160
http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22007.pdf
161

    
162

    
163
ARM-specific:
164

    
165
ARM Architecture Reference Manual (up to ARMv5TE):
166
http://www.arm.com/community/university/eulaarmarm.html
167

    
168
Procedure Call Standard for the ARM Architecture:
169
http://www.arm.com/pdfs/aapcs.pdf
170

    
171
Optimization guide for ARM9E (used in Nokia 770 Internet Tablet):
172
http://infocenter.arm.com/help/topic/com.arm.doc.ddi0240b/DDI0240A.pdf
173
Optimization guide for ARM11 (used in Nokia N800 Internet Tablet):
174
http://infocenter.arm.com/help/topic/com.arm.doc.ddi0211j/DDI0211J_arm1136_r1p5_trm.pdf
175
Optimization guide for Intel XScale (used in Sharp Zaurus PDA):
176
http://download.intel.com/design/intelxscale/27347302.pdf
177

    
178
PowerPC-specific:
179

    
180
PowerPC32/AltiVec PIM:
181
www.freescale.com/files/32bit/doc/ref_manual/ALTIVECPEM.pdf
182

    
183
PowerPC32/AltiVec PEM:
184
www.freescale.com/files/32bit/doc/ref_manual/ALTIVECPIM.pdf
185

    
186
CELL/SPU:
187
http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/30B3520C93F437AB87257060006FFE5E/$file/Language_Extensions_for_CBEA_2.4.pdf
188
http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/9F820A5FFA3ECE8C8725716A0062585F/$file/CBE_Handbook_v1.1_24APR2007_pub.pdf
189

    
190
SPARC-specific:
191
SPARC Joint Programming Specification (JPS1): Commonality
192
http://www.fujitsu.com/downloads/PRMPWR/JPS1-R1.0.4-Common-pub.pdf
193

    
194
UltraSPARC III Processor User's Manual (contains instruction timings)
195
http://www.sun.com/processors/manuals/USIIIv2.pdf
196

    
197
VIS Whitepaper (contains optimization guidelines)
198
http://www.sun.com/processors/vis/download/vis/vis_whitepaper.pdf
199

    
200
GCC asm links:
201
official doc but quite ugly
202
http://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html
203

    
204
a bit old (note "+" is valid for input-output, even though the next disagrees)
205
http://www.cs.virginia.edu/~clc5q/gcc-inline-asm.pdf