blob: c39e1e37b99f8fd39a3fb3d0b74228652df51bca
1 | optimization Tips (for libavcodec): |
2 | =================================== |
3 | |
4 | What to optimize: |
5 | ----------------- |
6 | If you plan to do non-x86 architecture specific optimizations (SIMD normally), |
7 | then take a look in the x86/ directory, as most important functions are |
8 | already optimized for MMX. |
9 | |
10 | If you want to do x86 optimizations then you can either try to fine-tune the |
11 | stuff in the x86 directory or find some other functions in the C source to |
12 | optimize, but there aren't many left. |
13 | |
14 | |
15 | Understanding these overoptimized functions: |
16 | -------------------------------------------- |
17 | As many functions tend to be a bit difficult to understand because |
18 | of optimizations, it can be hard to optimize them further, or write |
19 | architecture-specific versions. It is recommended to look at older |
20 | revisions of the interesting files (web frontends for the various FFmpeg |
21 | branches are listed at http://ffmpeg.org/download.html). |
22 | Alternatively, look into the other architecture-specific versions in |
23 | the x86/, ppc/, alpha/ subdirectories. Even if you don't exactly |
24 | comprehend the instructions, it could help understanding the functions |
25 | and how they can be optimized. |
26 | |
27 | NOTE: If you still don't understand some function, ask at our mailing list!!! |
28 | (http://lists.ffmpeg.org/mailman/listinfo/ffmpeg-devel) |
29 | |
30 | |
31 | When is an optimization justified? |
32 | ---------------------------------- |
33 | Normally, clean and simple optimizations for widely used codecs are |
34 | justified even if they only achieve an overall speedup of 0.1%. These |
35 | speedups accumulate and can make a big difference after awhile. Also, if |
36 | none of the following factors get worse due to an optimization -- speed, |
37 | binary code size, source size, source readability -- and at least one |
38 | factor improves, then an optimization is always a good idea even if the |
39 | overall gain is less than 0.1%. For obscure codecs that are not often |
40 | used, the goal is more toward keeping the code clean, small, and |
41 | readable instead of making it 1% faster. |
42 | |
43 | |
44 | WTF is that function good for ....: |
45 | ----------------------------------- |
46 | The primary purpose of this list is to avoid wasting time optimizing functions |
47 | which are rarely used. |
48 | |
49 | put(_no_rnd)_pixels{,_x2,_y2,_xy2} |
50 | Used in motion compensation (en/decoding). |
51 | |
52 | avg_pixels{,_x2,_y2,_xy2} |
53 | Used in motion compensation of B-frames. |
54 | These are less important than the put*pixels functions. |
55 | |
56 | avg_no_rnd_pixels* |
57 | unused |
58 | |
59 | pix_abs16x16{,_x2,_y2,_xy2} |
60 | Used in motion estimation (encoding) with SAD. |
61 | |
62 | pix_abs8x8{,_x2,_y2,_xy2} |
63 | Used in motion estimation (encoding) with SAD of MPEG-4 4MV only. |
64 | These are less important than the pix_abs16x16* functions. |
65 | |
66 | put_mspel8_mc* / wmv2_mspel8* |
67 | Used only in WMV2. |
68 | it is not recommended that you waste your time with these, as WMV2 |
69 | is an ugly and relatively useless codec. |
70 | |
71 | mpeg4_qpel* / *qpel_mc* |
72 | Used in MPEG-4 qpel motion compensation (encoding & decoding). |
73 | The qpel8 functions are used only for 4mv, |
74 | the avg_* functions are used only for B-frames. |
75 | Optimizing them should have a significant impact on qpel |
76 | encoding & decoding. |
77 | |
78 | qpel{8,16}_mc??_old_c / *pixels{8,16}_l4 |
79 | Just used to work around a bug in an old libavcodec encoder version. |
80 | Don't optimize them. |
81 | |
82 | add_bytes/diff_bytes |
83 | For huffyuv only, optimize if you want a faster ffhuffyuv codec. |
84 | |
85 | get_pixels / diff_pixels |
86 | Used for encoding, easy. |
87 | |
88 | clear_blocks |
89 | easiest to optimize |
90 | |
91 | gmc |
92 | Used for MPEG-4 gmc. |
93 | Optimizing this should have a significant effect on the gmc decoding |
94 | speed. |
95 | |
96 | gmc1 |
97 | Used for chroma blocks in MPEG-4 gmc with 1 warp point |
98 | (there are 4 luma & 2 chroma blocks per macroblock, so |
99 | only 1/3 of the gmc blocks use this, the other 2/3 |
100 | use the normal put_pixel* code, but only if there is |
101 | just 1 warp point). |
102 | Note: DivX5 gmc always uses just 1 warp point. |
103 | |
104 | pix_sum |
105 | Used for encoding. |
106 | |
107 | hadamard8_diff / sse / sad == pix_norm1 / dct_sad / quant_psnr / rd / bit |
108 | Specific compare functions used in encoding, it depends upon the |
109 | command line switches which of these are used. |
110 | Don't waste your time with dct_sad & quant_psnr, they aren't |
111 | really useful. |
112 | |
113 | put_pixels_clamped / add_pixels_clamped |
114 | Used for en/decoding in the IDCT, easy. |
115 | Note, some optimized IDCTs have the add/put clamped code included and |
116 | then put_pixels_clamped / add_pixels_clamped will be unused. |
117 | |
118 | idct/fdct |
119 | idct (encoding & decoding) |
120 | fdct (encoding) |
121 | difficult to optimize |
122 | |
123 | dct_quantize_trellis |
124 | Used for encoding with trellis quantization. |
125 | difficult to optimize |
126 | |
127 | dct_quantize |
128 | Used for encoding. |
129 | |
130 | dct_unquantize_mpeg1 |
131 | Used in MPEG-1 en/decoding. |
132 | |
133 | dct_unquantize_mpeg2 |
134 | Used in MPEG-2 en/decoding. |
135 | |
136 | dct_unquantize_h263 |
137 | Used in MPEG-4/H.263 en/decoding. |
138 | |
139 | |
140 | |
141 | Alignment: |
142 | Some instructions on some architectures have strict alignment restrictions, |
143 | for example most SSE/SSE2 instructions on x86. |
144 | The minimum guaranteed alignment is written in the .h files, for example: |
145 | void (*put_pixels_clamped)(const int16_t *block/*align 16*/, uint8_t *pixels/*align 8*/, ptrdiff_t stride); |
146 | |
147 | |
148 | General Tips: |
149 | ------------- |
150 | Use asm loops like: |
151 | __asm__( |
152 | "1: .... |
153 | ... |
154 | "jump_instruction .... |
155 | Do not use C loops: |
156 | do{ |
157 | __asm__( |
158 | ... |
159 | }while() |
160 | |
161 | For x86, mark registers that are clobbered in your asm. This means both |
162 | general x86 registers (e.g. eax) as well as XMM registers. This last one is |
163 | particularly important on Win64, where xmm6-15 are callee-save, and not |
164 | restoring their contents leads to undefined results. In external asm (e.g. |
165 | yasm), you do this by using: |
166 | cglobal function_name, num_args, num_regs, num_xmm_regs |
167 | In inline asm, you specify clobbered registers at the end of your asm: |
168 | __asm__(".." ::: "%eax"). |
169 | If gcc is not set to support sse (-msse) it will not accept xmm registers |
170 | in the clobber list. For that we use two macros to declare the clobbers. |
171 | XMM_CLOBBERS should be used when there are other clobbers, for example: |
172 | __asm__(".." ::: XMM_CLOBBERS("xmm0",) "eax"); |
173 | and XMM_CLOBBERS_ONLY should be used when the only clobbers are xmm registers: |
174 | __asm__(".." :: XMM_CLOBBERS_ONLY("xmm0")); |
175 | |
176 | Do not expect a compiler to maintain values in your registers between separate |
177 | (inline) asm code blocks. It is not required to. For example, this is bad: |
178 | __asm__("movdqa %0, %%xmm7" : src); |
179 | /* do something */ |
180 | __asm__("movdqa %%xmm7, %1" : dst); |
181 | - first of all, you're assuming that the compiler will not use xmm7 in |
182 | between the two asm blocks. It probably won't when you test it, but it's |
183 | a poor assumption that will break at some point for some --cpu compiler flag |
184 | - secondly, you didn't mark xmm7 as clobbered. If you did, the compiler would |
185 | have restored the original value of xmm7 after the first asm block, thus |
186 | rendering the combination of the two blocks of code invalid |
187 | Code that depends on data in registries being untouched, should be written as |
188 | a single __asm__() statement. Ideally, a single function contains only one |
189 | __asm__() block. |
190 | |
191 | Use external asm (nasm/yasm) or inline asm (__asm__()), do not use intrinsics. |
192 | The latter requires a good optimizing compiler which gcc is not. |
193 | |
194 | When debugging a x86 external asm compilation issue, if lost in the macro |
195 | expansions, add DBG=1 to your make command-line: the input file will be |
196 | preprocessed, stripped of the debug/empty lines, then compiled, showing the |
197 | actual lines causing issues. |
198 | |
199 | Inline asm vs. external asm |
200 | --------------------------- |
201 | Both inline asm (__asm__("..") in a .c file, handled by a compiler such as gcc) |
202 | and external asm (.s or .asm files, handled by an assembler such as yasm/nasm) |
203 | are accepted in FFmpeg. Which one to use differs per specific case. |
204 | |
205 | - if your code is intended to be inlined in a C function, inline asm is always |
206 | better, because external asm cannot be inlined |
207 | - if your code calls external functions, yasm is always better |
208 | - if your code takes huge and complex structs as function arguments (e.g. |
209 | MpegEncContext; note that this is not ideal and is discouraged if there |
210 | are alternatives), then inline asm is always better, because predicting |
211 | member offsets in complex structs is almost impossible. It's safest to let |
212 | the compiler take care of that |
213 | - in many cases, both can be used and it just depends on the preference of the |
214 | person writing the asm. For new asm, the choice is up to you. For existing |
215 | asm, you'll likely want to maintain whatever form it is currently in unless |
216 | there is a good reason to change it. |
217 | - if, for some reason, you believe that a particular chunk of existing external |
218 | asm could be improved upon further if written in inline asm (or the other |
219 | way around), then please make the move from external asm <-> inline asm a |
220 | separate patch before your patches that actually improve the asm. |
221 | |
222 | |
223 | Links: |
224 | ====== |
225 | http://www.aggregate.org/MAGIC/ |
226 | |
227 | x86-specific: |
228 | ------------- |
229 | http://developer.intel.com/design/pentium4/manuals/248966.htm |
230 | |
231 | The IA-32 Intel Architecture Software Developer's Manual, Volume 2: |
232 | Instruction Set Reference |
233 | http://developer.intel.com/design/pentium4/manuals/245471.htm |
234 | |
235 | http://www.agner.org/assem/ |
236 | |
237 | AMD Athlon Processor x86 Code Optimization Guide: |
238 | http://www.amd.com/us-en/assets/content_type/white_papers_and_tech_docs/22007.pdf |
239 | |
240 | |
241 | ARM-specific: |
242 | ------------- |
243 | ARM Architecture Reference Manual (up to ARMv5TE): |
244 | http://www.arm.com/community/university/eulaarmarm.html |
245 | |
246 | Procedure Call Standard for the ARM Architecture: |
247 | http://www.arm.com/pdfs/aapcs.pdf |
248 | |
249 | Optimization guide for ARM9E (used in Nokia 770 Internet Tablet): |
250 | http://infocenter.arm.com/help/topic/com.arm.doc.ddi0240b/DDI0240A.pdf |
251 | Optimization guide for ARM11 (used in Nokia N800 Internet Tablet): |
252 | http://infocenter.arm.com/help/topic/com.arm.doc.ddi0211j/DDI0211J_arm1136_r1p5_trm.pdf |
253 | Optimization guide for Intel XScale (used in Sharp Zaurus PDA): |
254 | http://download.intel.com/design/intelxscale/27347302.pdf |
255 | Intel Wireless MMX 2 Coprocessor: Programmers Reference Manual |
256 | http://download.intel.com/design/intelxscale/31451001.pdf |
257 | |
258 | PowerPC-specific: |
259 | ----------------- |
260 | PowerPC32/AltiVec PIM: |
261 | www.freescale.com/files/32bit/doc/ref_manual/ALTIVECPEM.pdf |
262 | |
263 | PowerPC32/AltiVec PEM: |
264 | www.freescale.com/files/32bit/doc/ref_manual/ALTIVECPIM.pdf |
265 | |
266 | CELL/SPU: |
267 | http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/30B3520C93F437AB87257060006FFE5E/$file/Language_Extensions_for_CBEA_2.4.pdf |
268 | http://www-01.ibm.com/chips/techlib/techlib.nsf/techdocs/9F820A5FFA3ECE8C8725716A0062585F/$file/CBE_Handbook_v1.1_24APR2007_pub.pdf |
269 | |
270 | GCC asm links: |
271 | -------------- |
272 | official doc but quite ugly |
273 | http://gcc.gnu.org/onlinedocs/gcc/Extended-Asm.html |
274 | |
275 | a bit old (note "+" is valid for input-output, even though the next disagrees) |
276 | http://www.cs.virginia.edu/~clc5q/gcc-inline-asm.pdf |
277 |