Recherche avancée

Sur d’autres sites (7642)

• Announcing the world’s fastest VP8 decoder : ffvp8

  24 juillet 2010, par Dark Shikari — VP8, ffmpeg, google, speed

  Back when I originally reviewed VP8, I noted that the official decoder, libvpx, was rather slow. While there was no particular reason that it should be much faster than a good H.264 decoder, it shouldn’t have been that much slower either ! So, I set out with Ronald Bultje and David Conrad to make a better one in FFmpeg. This one would be community-developed and free from the beginning, rather than the proprietary code-dump that was libvpx. A few weeks ago the decoder was complete enough to be bit-exact with libvpx, making it the first independent free implementation of a VP8 decoder. Now, with the first round of optimizations complete, it should be ready for primetime. I’ll go into some detail about the development process, but first, let’s get to the real meat of this post : the benchmarks.

  We tested on two 1080p clips : Parkjoy, a live-action 1080p clip, and the Sintel trailer, a CGI 1080p clip. Testing was done using “time ffmpeg -vcodec libvpx or vp8 -i input -vsync 0 -an -f null -”. We all used the latest SVN FFmpeg at the time of this posting ; the last revision optimizing the VP8 decoder was r24471.

  

  As these benchmarks show, ffvp8 is clearly much faster than libvpx, particularly on 64-bit. It’s even faster by a large margin on Atom, despite the fact that we haven’t even begun optimizing for it. In many cases, ffvp8′s extra speed can make the difference between a video that plays and one that doesn’t, especially in modern browsers with software compositing engines taking up a lot of CPU time. Want to get faster playback of VP8 videos ? The next versions of FFmpeg-based players, like VLC, will include ffvp8. Want to get faster playback of WebM in your browser ? Lobby your browser developers to use ffvp8 instead of libvpx. I expect Chrome to switch first, as they already use libavcodec for most of their playback system.

  Keep in mind ffvp8 is not “done” — we will continue to improve it and make it faster. We still have a number of optimizations in the pipeline that aren’t committed yet.

  Developing ffvp8

  The initial challenge, primarily pioneered by David and Ronald, was constructing the core decoder and making it bit-exact to libvpx. This was rather challenging, especially given the lack of a real spec. Many parts of the spec were outright misleading and contradicted libvpx itself. It didn’t help that the suite of official conformance tests didn’t even cover all the features used by the official encoder ! We’ve already started adding our own conformance tests to deal with this. But I’ve complained enough in past posts about the lack of a spec ; let’s get onto the gritty details.

  The next step was adding SIMD assembly for all of the important DSP functions. VP8′s motion compensation and deblocking filter are by far the most CPU-intensive parts, much the same as in H.264. Unlike H.264, the deblocking filter relies on a lot of internal saturation steps, which are free in SIMD but costly in a normal C implementation, making the plain C code even slower. Of course, none of this is a particularly large problem ; any sane video decoder has all this stuff in SIMD.

  I tutored Ronald in x86 SIMD and wrote most of the motion compensation, intra prediction, and some inverse transforms. Ronald wrote the rest of the inverse transforms and a bit of the motion compensation. He also did the most difficult part : the deblocking filter. Deblocking filters are always a bit difficult because every one is different. Motion compensation, by comparison, is usually very similar regardless of video format ; a 6-tap filter is a 6-tap filter, and most of the variation going on is just the choice of numbers to multiply by.

  The biggest challenge in an SIMD deblocking filter is to avoid unpacking, that is, going from 8-bit to 16-bit. Many operations in deblocking filters would naively appear to require more than 8-bit precision. A simple example in the case of x86 is abs(a-b), where a and b are 8-bit unsigned integers. The result of “a-b” requires a 9-bit signed integer (it can be anywhere from -255 to 255), so it can’t fit in 8-bit. But this is quite possible to do without unpacking : (satsub(a,b) | satsub(b,a)), where “satsub” performs a saturating subtract on the two values. If the value is positive, it yields the result ; if the value is negative, it yields zero. Oring the two together yields the desired result. This requires 4 ops on x86 ; unpacking would probably require at least 10, including the unpack and pack steps.

  After the SIMD came optimizing the C code, which still took a significant portion of the total runtime. One of my biggest optimizations was adding aggressive “smart” prefetching to reduce cache misses. ffvp8 prefetches the reference frames (PREVIOUS, GOLDEN, and ALTREF)… but only the ones which have been used reasonably often this frame. This lets us prefetch everything we need without prefetching things that we probably won’t use. libvpx very often encodes frames that almost never (but not quite never) use GOLDEN or ALTREF, so this optimization greatly reduces time spent prefetching in a lot of real videos. There are of course countless other optimizations we made that are too long to list here as well, such as David’s entropy decoder optimizations. I’d also like to thank Eli Friedman for his invaluable help in benchmarking a lot of these changes.

  What next ? Altivec (PPC) assembly is almost nonexistent, with the only functions being David’s motion compensation code. NEON (ARM) is completely nonexistent : we’ll need that to be fast on mobile devices as well. Of course, all this will come in due time — and as always — patches welcome !

  Appendix : the raw numbers

  Here’s the raw numbers (in fps) for the graphs at the start of this post, with standard error values :

  Core i7 620QM (1.6Ghz), Windows 7, 32-bit :
  Parkjoy ffvp8 : 44.58 0.44
  Parkjoy libvpx : 33.06 0.23
  Sintel ffvp8 : 74.26 1.18
  Sintel libvpx : 56.11 0.96

  Core i5 520M (2.4Ghz), Linux, 64-bit :
  Parkjoy ffvp8 : 68.29 0.06
  Parkjoy libvpx : 41.06 0.04
  Sintel ffvp8 : 112.38 0.37
  Sintel libvpx : 69.64 0.09

  Core 2 T9300 (2.5Ghz), Mac OS X 10.6.4, 64-bit :
  Parkjoy ffvp8 : 54.09 0.02
  Parkjoy libvpx : 33.68 0.01
  Sintel ffvp8 : 87.54 0.03
  Sintel libvpx : 52.74 0.04

  Core Duo (2Ghz), Mac OS X 10.6.4, 32-bit :
  Parkjoy ffvp8 : 21.31 0.02
  Parkjoy libvpx : 17.96 0.00
  Sintel ffvp8 : 41.24 0.01
  Sintel libvpx : 29.65 0.02

  Atom N270 (1.6Ghz), Linux, 32-bit :
  Parkjoy ffvp8 : 15.29 0.01
  Parkjoy libvpx : 12.46 0.01
  Sintel ffvp8 : 26.87 0.05
  Sintel libvpx : 20.41 0.02

• Beware the builtins

  14 janvier 2010, par Mans — Compilers

  GCC includes a large number of builtin functions allegedly providing optimised code for common operations not easily expressed directly in C. Rather than taking such claims at face value (this is GCC after all), I decided to conduct a small investigation to see how well a few of these functions are actually implemented for various targets.

  For my test, I selected the following functions :

  • __builtin_bswap32 : Byte-swap a 32-bit word.
  • __builtin_bswap64 : Byte-swap a 64-bit word.
  • __builtin_clz : Count leading zeros in a word.
  • __builtin_ctz : Count trailing zeros in a word.
  • __builtin_prefetch : Prefetch data into cache.

  To test the quality of these builtins, I wrapped each in a normal function, then compiled the code for these targets :

  • ARMv7
  • AVR32
  • MIPS
  • MIPS64
  • PowerPC
  • PowerPC64
  • x86
  • x86_64

  In all cases I used compiler flags were -O3 -fomit-frame-pointer plus any flags required to select a modern CPU model.
  

  ARM

  Both __builtin_clz and __builtin_prefetch generate the expected CLZ and PLD instructions respectively. The code for __builtin_ctz is reasonable for ARMv6 and earlier :

  rsb     r3, r0, #0
  and     r0, r3, r0
  clz     r0, r0
  rsb     r0, r0, #31
  

  For ARMv7 (in fact v6T2), however, using the new bit-reversal instruction would have been better :

  rbit    r0, r0
  clz     r0, r0
  

  I suspect this is simply a matter of the function not yet having been updated for ARMv7, which is perhaps even excusable given the relatively rare use cases for it.

  The byte-reversal functions are where it gets shocking. Rather than use the REV instruction found from ARMv6 on, both of them generate external calls to __bswapsi2 and __bswapdi2 in libgcc, which is plain C code :

  SItype
  __bswapsi2 (SItype u)
  
    return ((((u) & 0xff000000) >> 24)
            | (((u) & 0x00ff0000) >>  8)
            | (((u) & 0x0000ff00) <<  8)
            | (((u) & 0x000000ff) << 24)) ;
  
  DItype
  
  __bswapdi2 (DItype u)
  
  
  
     return ((((u) & 0xff00000000000000ull) >> 56)
  
            | (((u) & 0x00ff000000000000ull) >> 40)
  
            | (((u) & 0x0000ff0000000000ull) >> 24)
  
            | (((u) & 0x000000ff00000000ull) >>  8)
  
            | (((u) & 0x00000000ff000000ull) <<  8)
  
            | (((u) & 0x0000000000ff0000ull) << 24)
  
            | (((u) & 0x000000000000ff00ull) << 40)
  
            | (((u) & 0x00000000000000ffull) << 56)) ;
  
  
  

  While the 32-bit version compiles to a reasonable-looking shift/mask/or job, the 64-bit one is a real WTF. Brace yourselves :

  push    r4, r5, r6, r7, r8, r9, sl, fp
  mov     r5, #0
  mov     r6, #65280 ; 0xff00
  sub     sp, sp, #40 ; 0x28
  and     r7, r0, r5
  and     r8, r1, r6
  str     r7, [sp, #8]
  str     r8, [sp, #12]
  mov     r9, #0
  mov     r4, r1
  and     r5, r0, r9
  mov     sl, #255 ; 0xff
  ldr     r9, [sp, #8]
  and     r6, r4, sl
  mov     ip, #16711680 ; 0xff0000
  str     r5, [sp, #16]
  str     r6, [sp, #20]
  lsl     r2, r0, #24
  and     ip, ip, r1
  lsr     r7, r4, #24
  mov     r1, #0
  lsr     r5, r9, #24
  mov     sl, #0
  mov     r9, #-16777216 ; 0xff000000
  and     fp, r0, r9
  lsr     r6, ip, #8
  orr     r9, r7, r1
  and     ip, r4, sl
  orr     sl, r1, r2
  str     r6, [sp]
  str     r9, [sp, #32]
  str     sl, [sp, #36] ; 0x24
  add     r8, sp, #32
  ldm     r8, r7, r8
  str     r1, [sp, #4]
  ldm     sp, r9, sl
  orr     r7, r7, r9
  orr     r8, r8, sl
  str     r7, [sp, #32]
  str     r8, [sp, #36] ; 0x24
  mov     r3, r0
  mov     r7, #16711680 ; 0xff0000
  mov     r8, #0
  and     r9, r3, r7
  and     sl, r4, r8
  ldr     r0, [sp, #16]
  str     fp, [sp, #24]
  str     ip, [sp, #28]
  stm     sp, r9, sl
  ldr     r7, [sp, #20]
  ldr     sl, [sp, #12]
  ldr     fp, [sp, #12]
  ldr     r8, [sp, #28]
  lsr     r0, r0, #8
  orr     r7, r0, r7, lsl #24
  lsr     r6, sl, #24
  orr     r5, r5, fp, lsl #8
  lsl     sl, r8, #8
  mov     fp, r7
  add     r8, sp, #32
  ldm     r8, r7, r8
  orr     r6, r6, r8
  ldr     r8, [sp, #20]
  ldr     r0, [sp, #24]
  orr     r5, r5, r7
  lsr     r8, r8, #8
  orr     sl, sl, r0, lsr #24
  mov     ip, r8
  ldr     r0, [sp, #4]
  orr     fp, fp, r5
  ldr     r5, [sp, #24]
  orr     ip, ip, r6
  ldr     r6, [sp]
  lsl     r9, r5, #8
  lsl     r8, r0, #24
  orr     fp, fp, r9
  lsl     r3, r3, #8
  orr     r8, r8, r6, lsr #8
  orr     ip, ip, sl
  lsl     r7, r6, #24
  and     r5, r3, #16711680 ; 0xff0000
  orr     r7, r7, fp
  orr     r8, r8, ip
  orr     r4, r1, r7
  orr     r5, r5, r8
  mov     r9, r6
  mov     r1, r5
  mov     r0, r4
  add     sp, sp, #40 ; 0x28
  pop     r4, r5, r6, r7, r8, r9, sl, fp
  bx      lr
  

  That’s right, 91 instructions to move 8 bytes around a bit. GCC definitely has a problem with 64-bit numbers. It is perhaps worth noting that the bswap_64 macro in glibc splits the 64-bit value into 32-bit halves which are then reversed independently, thus side-stepping this weakness of gcc.

  As a side note, ARM RVCT (armcc) compiles those functions perfectly into one and two REV instructions, respectively.

  AVR32

  There is not much to report here. The latest gcc version available is 4.2.4, which doesn’t appear to have the bswap functions. The other three are handled nicely, even using a bit-reverse for __builtin_ctz.

  MIPS / MIPS64

  The situation MIPS is similar to ARM. Both bswap builtins result in external libgcc calls, the rest giving sensible code.

  PowerPC

  I scarcely believe my eyes, but this one is actually not bad. The PowerPC has no byte-reversal instructions, yet someone seems to have taken the time to teach gcc a good instruction sequence for this operation. The PowerPC does have some powerful rotate-and-mask instructions which come in handy here. First the 32-bit version :

  rotlwi  r0,r3,8
  rlwimi  r0,r3,24,0,7
  rlwimi  r0,r3,24,16,23
  mr      r3,r0
  blr
  

  The 64-bit byte-reversal simply applies the above code on each half of the value :

  rotlwi  r0,r3,8
  rlwimi  r0,r3,24,0,7
  rlwimi  r0,r3,24,16,23
  rotlwi  r3,r4,8
  rlwimi  r3,r4,24,0,7
  rlwimi  r3,r4,24,16,23
  mr      r4,r0
  blr
  

  Although I haven’t analysed that code carefully, it looks pretty good.

  PowerPC64

  Doing 64-bit operations is easier on a 64-bit CPU, right ? For you and me perhaps, but not for gcc. Here __builtin_bswap64 gives us the now familiar __bswapdi2 call, and while not as bad as the ARM version, it is not pretty :

  rldicr  r0,r3,8,55
  rldicr  r10,r3,56,7
  rldicr  r0,r0,56,15
  rldicl  r11,r3,8,56
  rldicr  r9,r3,16,47
  or      r11,r10,r11
  rldicr  r9,r9,48,23
  rldicl  r10,r0,24,40
  rldicr  r0,r3,24,39
  or      r11,r11,r10
  rldicl  r9,r9,40,24
  rldicr  r0,r0,40,31
  or      r9,r11,r9
  rlwinm  r10,r3,0,0,7
  rldicl  r0,r0,56,8
  or      r0,r9,r0
  rldicr  r10,r10,8,55
  rlwinm  r11,r3,0,8,15
  or      r0,r0,r10
  rldicr  r11,r11,24,39
  rlwinm  r3,r3,0,16,23
  or      r0,r0,r11
  rldicr  r3,r3,40,23
  or      r3,r0,r3
  blr
  

  That is 6 times longer than the (presumably) hand-written 32-bit version.

  x86 / x86_64

  As one might expect, results on x86 are good. All the tested functions use the available special instructions. One word of caution though : the bit-counting instructions are very slow on some implementations, specifically the Atom, AMD chips, and the notoriously slow Pentium4E.

  Conclusion

  In conclusion, I would say gcc builtins can be useful to avoid fragile inline assembler. Before using them, however, one should make sure they are not in fact harmful on the required targets. Not even those builtins mapping directly to CPU instructions can be trusted.

• The first in-depth technical analysis of VP8

    
  19 mai 2010, par Dark Shikari — VP8, google

  Back in my original post about Internet video, I made some initial comments on the hope that VP8 would solve the problems of web video by providing a supposed patent-free video format with significantly better compression than the current options of Theora and Dirac. Fortunately, it seems I was able to acquire access to the VP8 spec, software, and source a good few days before the official release and so was able to perform a detailed technical analysis in time for the official release.

  The questions I will try to answer here are :

  1. How good is VP8 ? Is the file format actually better than H.264 in terms of compression, and could a good VP8 encoder beat x264 ? On2 claimed 50% better than H.264, but On2 has always made absurd claims that they were never able to back up with results, so such a number is almost surely wrong. VP7, for example, was claimed to be 15% better than H.264 while being much faster, but was in reality neither faster nor higher quality.

  2. How good is On2′s VP8 implementation ? Irrespective of how good the spec is, is the implementation good, or is this going to be just like VP3, where On2 releases an unusably bad implementation with the hope that the community will fix it for them ? Let’s hope not ; it took 6 years to fix Theora !

  3. How likely is VP8 to actually be free of patents ? Even if VP8 is worse than H.264, being patent-free is still a useful attribute for obvious reasons. But as noted in my previous post, merely being published by Google doesn’t guarantee that it is. Microsoft did similar a few years ago with the release of VC-1, which was claimed to be patent-free — but within mere months after release, a whole bunch of companies claimed patents on it and soon enough a patent pool was formed.

  We’ll start by going through the core features of VP8. We’ll primarily analyze them by comparing to existing video formats. Keep in mind that an encoder and a spec are two different things : it’s possible for good encoder to be written for a bad spec or vice versa ! Hence why a really good MPEG-1 encoder can beat a horrific H.264 encoder.

  But first, a comment on the spec itself.

  AAAAAAAGGGGGGGGGGGGGHHHHHHHHHHHHHHHHHHHHHHHHHHHHHH !

  The spec consists largely of C code copy-pasted from the VP8 source code — up to and including TODOs, “optimizations”, and even C-specific hacks, such as workarounds for the undefined behavior of signed right shift on negative numbers. In many places it is simply outright opaque. Copy-pasted C code is not a spec. I may have complained about the H.264 spec being overly verbose, but at least it’s precise. The VP8 spec, by comparison, is imprecise, unclear, and overly short, leaving many portions of the format very vaguely explained. Some parts even explicitly refuse to fully explain a particular feature, pointing to highly-optimized, nigh-impossible-to-understand reference code for an explanation. There’s no way in hell anyone could write a decoder solely with this spec alone.

  Now that I’ve gotten that out of my system, let’s get back to VP8 itself. To begin with, to get a general sense for where all this fits in, basically all modern video formats work via some variation on the following chain of steps :

  Encode : Predict -> Transform + Quant -> Entropy Code -> Loopfilter
  Decode : Entropy Decode -> Predict -> Dequant + Inverse Transform -> Loopfilter

  If you’re looking to just get to the results and skip the gritty technical details, make sure to check out the “overall verdict” section and the “visual results” section. Or at least skip to the “summary for the lazy”.
  

  Prediction

  Prediction is any step which attempts to guess the content of an area of the frame. This could include functions based on already-known pixels in the same frame (e.g. inpainting) or motion compensation from a previous frame. Prediction usually involves side data, such as a signal telling the decoder a motion vector to use for said motion compensation.

  Intra Prediction

  Intra prediction is used to guess the content of a block without referring to other frames. VP8′s intra prediction is basically ripped off wholesale from H.264 : the “subblock” prediction modes are almost exactly identical (they even have the same names !) to H.264′s i4x4 mode, and the whole block prediction mode is basically identical to i16x16. Chroma prediction modes are practically identical as well. i8x8, from H.264 High Profile, is not present. An additional difference is that the planar prediction mode has been replaced with TM_PRED, a very vaguely similar analogue. The specific prediction modes are internally slightly different, but have the same names as in H.264.

  Honestly, I’m very disappointed here. While H.264′s intra prediction is good, it has certainly been improved on quite a bit over the past 7 years, and I thought that blatantly ripping it off was the domain of companies like Real (see RV40). I expected at least something slightly more creative out of On2. But more important than any of that : this is a patent time-bomb waiting to happen. H.264′s spatial intra prediction is covered in patents and I don’t think that On2 will be able to just get away with changing the rounding in the prediction modes. I’d like to see Google’s justification for this — they must have a good explanation for why they think there won’t be any patent issues.

  Update : spatial intra prediction apparently dates back to Nokia’s MVC H.26L proposal, from around 2000. It’s possible that Google believes that this is sufficient prior art to invalidate existing patents — which is not at all unreasonable !

  Verdict on Intra Prediction : Slightly modified ripoff of H.264. Somewhat worse than H.264 due to omission of i8x8.

  Inter Prediction

  Inter prediction is used to guess the content of a block by referring to past frames. There are two primary components to inter prediction : reference frames and motion vectors. The reference frame is a past frame from which to grab pixels from and the motion vectors index an offset into that frame. VP8 supports a total of 3 reference frames : the previous frame, the “alt ref” frame, and the “golden frame”. For motion vectors, VP8 supports variable-size partitions much like H.264. For subpixel precision, it supports quarter-pel motion vectors with a 6-tap interpolation filter. In short :

  VP8 reference frames : up to 3
  H.264 reference frames : up to 16
  VP8 partition types : 16×16, 16×8, 8×16, 8×8, 4×4
  H.264 partition types : 16×16, 16×8, 8×16, flexible subpartitions (each 8×8 can be 8×8, 8×4, 4×8, or 4×4).
  VP8 chroma MV derivation : each 4×4 chroma block uses the average of colocated luma MVs (same as MPEG-4 ASP)
  H.264 chroma MV derivation : chroma uses luma MVs directly
  VP8 interpolation filter : qpel, 6-tap luma, mixed 4/6-tap chroma
  H.264 interpolation filter : qpel, 6-tap luma (staged filter), bilinear chroma
  H.264 has but VP8 doesn’t : B-frames, weighted prediction

  H.264 has a significantly better and more flexible referencing structure. Sub-8×8 partitions are mostly unnecessary, so VP8′s omission of the H.264-style subpartitions has little consequence. The chroma MV derivation is more accurate in H.264 but slightly slower ; in practice the difference is probably near-zero both speed and compression-wise, since sub-8×8 luma partitions are rarely used (and I would suspect the same carries over to VP8).

  The VP8 interpolation filter is likely slightly better, but will definitely be slower to implement, both encoder and decoder-side. A staged filter allows the encoder to precalculate all possible halfpel positions and then quickly calculate qpel positions when necessary : an unstaged filter does not, making subpel motion estimation much slower. Not that unstaged filters are bad — staged filters have basically been abandoned for all of the H.265 proposals — it’s just an inherent disadvantage performance-wise. Additionally, having as high as 6 taps on chroma is, IMO, completely unnecessary and wasteful.

  The lack of B-frames in VP8 is a killer. B-frames can give 10-20% (or more) compression benefit for minimal speed cost ; their omission in VP8 probably costs more compression than all other problems noted in this post combined. This was not unexpected, however ; On2 has never used B-frames in any of their video formats. They also likely present serious patent problems, which probably explains their omission. Lack of weighted prediction is also going to hurt a bit, especially in fades.

  Update : Alt-ref frames can apparently be used to partially replicate the lack of B-frames. It’s not nearly as good, but it can get at least some of the benefit without actual B-frames.

  Verdict on Inter Prediction : Similar partitioning structure to H.264. Much weaker referencing structure. More complex, slightly better interpolation filter. Mostly a wash — except for the lack of B-frames, which is seriously going to hurt compression.
  

  Transform and Quantization
  

  After prediction, the encoder takes the difference between the prediction and the actual source pixels (the residual), transforms it, and quantizes it. The transform step is designed to make the data more amenable to compression by decorrelating it. The quantization step is the actual information-losing step where compression occurs ; the output values of transform are rounded, mostly to zero, leaving only a few integer coefficients.

  Transform

  For transform, VP8 uses again a very H.264-reminiscent scheme. Each 16×16 macroblock is divided into 16 4×4 DCT blocks, each of which is transformed by a bit-exact DCT approximation. Then, the DC coefficients of each block are collected into another 4×4 group, which is then Hadamard-transformed. OK, so this isn’t reminiscent of H.264, this is H.264. There are, however, 3 differences between VP8′s scheme and H.264′s.

  The first is that the 8×8 transform is omitted entirely (fitting with the omission of the i8x8 intra mode). The second is the specifics of the transform itself. H.264 uses an extremely simplified “DCT” which is so un-DCT-like that it often referred to as the HCT (H.264 Cosine Transform) instead. This simplified transform results in roughly 1% worse compression, but greatly simplifies the transform itself, which can be implemented entirely with adds, subtracts, and right shifts by 1. VC-1 uses a more accurate version that relies on a few small multiplies (numbers like 17, 22, 10, etc). VP8 uses an extremely, needlessly accurate version that uses very large multiplies (20091 and 35468). This in retrospect is not surpising, as it is very similar to what VP3 used.

  The third difference is that the Hadamard hierarchical transform is applied for some inter blocks, not merely i16x16. In particular, it also runs for p16x16 blocks. While this is definitely a good idea, especially given the small transform size (and the need to decorrelate the DC value between the small transforms), I’m not quite sure I agree with the decision to limit it to p16x16 blocks ; it seems that perhaps with a small amount of modification this could also be useful for other motion partitions. Also, note that unlike H.264, the hierarchical transform is luma-only and not applied to chroma.

  Overall, the transform scheme in VP8 is definitely weaker than in H.264. The lack of an 8×8 transform is going to have a significant impact on detail retention, especially at high resolutions. The transform is needlessly slower than necessary as well, though a shift-based transform might be out of the question due to patents. The one good new idea here is applying the hierarchical DC transform to inter blocks.

  Verdict on Transform : Similar to H.264. Slower, slightly more accurate 4×4 transform. Improved DC transform for luma (but not on chroma). No 8×8 transform. Overall, worse.

  Quantization

  For quantization, the core process is basically the same among all MPEG-like video formats, and VP8 is no exception. The primary ways that video formats tend to differentiate themselves here is by varying quantization scaling factors. There are two ways in which this is primarily done : frame-based offsets that apply to all coefficients or just some portion of them, and macroblock-level offsets. VP8 primarily uses the former ; in a scheme much less flexible than H.264′s custom quantization matrices, it allows for adjusting the quantizer of luma DC, luma AC, chroma DC, and so forth, separately. The latter (macroblock-level quantizer choice) can, in theory, be done using its “segmentation map” features, albeit very hackily and not very efficiently.

  The killer mistake that VP8 has made here is not making macroblock-level quantization a core feature of VP8. Algorithms that take advantage of macroblock-level quantization are known as “adaptive quantization” and are absolutely critical to competitive visual quality. My implementation of variance-based adaptive quantization (before, after) in x264 still stands to this day as the single largest visual quality gain in x264 history. Encoder comparisons have showed over and over that encoders without adaptive quantization simply cannot compete.

  Thus, while adaptive quantization is possible in VP8, the only way to implement it is to define one segment map for every single quantizer that one wants and to code the segment map index for every macroblock. This is inefficient and cumbersome ; even the relatively suboptimal MPEG-style delta quantizer system would be a better option. Furthermore, only 4 segment maps are allowed, for a maximum of 4 quantizers per frame.

  Verdict on Quantization : Lack of well-integrated adaptive quantization is going to be a killer when the time comes to implement psy optimizations. Overall, much worse.

  Entropy Coding

  Entropy coding is the process of taking all the information from all the other processes : DCT coefficients, prediction modes, motion vectors, and so forth — and compressing them losslessly into the final output file. VP8 uses an arithmetic coder somewhat similar to H.264′s, but with a few critical differences. First, it omits the range/probability table in favor of a multiplication. Second, it is entirely non-adaptive : unlike H.264′s, which adapts after every bit decoded, probability values are constant over the course of the frame. Accordingly, the encoder may periodically send updated probability values in frame headers for some syntax elements. Keyframes reset the probability values to the defaults.

  This approach isn’t surprising ; VP5 and VP6 (and probably VP7) also used non-adaptive arithmetic coders. How much of a penalty this actually means compression-wise is unknown ; it’s not easy to measure given the design of either H.264 or VP8. More importantly, I question the reason for this : making it adaptive would add just one single table lookup to the arithmetic decoding function — hardly a very large performance impact.

  Of course, the arithmetic coder is not the only part of entropy coding : an arithmetic coder merely turns 0s and 1s into an output bitstream. The process of creating those 0s and 1s and selecting the probabilities for the encoder to use is an equally interesting problem. Since this is a very complicated part of the video format, I’ll just comment on the parts that I found particularly notable.

  Motion vector coding consists of two parts : prediction based on neighboring motion vectors and the actual compression of the resulting delta between that and the actual motion vector. The prediction scheme in VP8 is a bit odd — worse, the section of the spec covering this contains no English explanation, just confusingly-written C code. As far as I can tell, it chooses an arithmetic coding context based on the neighboring MVs, then decides which of the predicted motion vectors to use, or whether to code a delta instead.

  The downside of this scheme is that, like in VP3/Theora (though not nearly as badly), it biases heavily towards the re-use of previous motion vectors. This is dangerous because, as the Theora devs have recently found (and fixed to some extent in Theora 1.2 aka Ptalabvorm), any situation in which the encoder picks a motion vector which isn’t the “real” motion vector in order to save bits can potentially have negative visual consequences. In terms of raw efficiency, I’m not sure whether VP8 or H.264′s prediction is better here.

  The compression of the resulting delta is similar to H.264, except for the coding of very large deltas, which is slightly better (similar to FFV1′s Golomb-like arithmetic codes).

  Intra prediction mode coding is done using arithmetic coding contexts based on the modes of the neighboring blocks. This is probably a good bit better than the hackneyed method that H.264 uses, which always struck me as being poorly designed.

  Residual coding is even more difficult to understand than motion vector coding, as the only full reference is a bunch of highly optimized, highly obfuscated C code. Like H.264′s CAVLC, it bases contexts on the number of nonzero coefficients in the top and left blocks relative to the current block. In addition, it also considers the magnitude of those coefficients and, like H.264′s CABAC, updates as coefficients are decoded.

  One more thing to note is the data partitioning scheme used by VP8. This scheme is much like VP3/Theora’s and involves putting each syntax element in its own component of the bitstream. The unfortunate problem with this is that it’s a nightmare for hardware implementations, greatly increasing memory bandwidth requirements. I have already received a complaint from a hardware developer about this specific feature with regard to VP8.

  Verdict on Entropy Coding : I’m not quite sure here. It’s better in some ways, worse in some ways, and just plain weird in others. My hunch is that it’s probably a very slight win for H.264 ; non-adaptive arithmetic coding has to have some serious penalties. It may also be a hardware implementation problem.
  

  Loop Filter

  The loop filter is run after decoding or encoding a frame and serves to perform extra processing on a frame, usually to remove blockiness in DCT-based video formats. Unlike postprocessing, this is not only for visual reasons, but also to improve prediction for future frames. Thus, it has to be done identically in both the encoder and decoder. VP8′s loop filter is vaguely similar to H.264′s, but with a few differences. First, it has two modes (which can be chosen by the encoder) : a fast mode and a normal mode. The fast mode is somewhat simpler than H.264′s, while the normal mode is somewhat more complex. Secondly, when filtering between macroblocks, VP8′s filter has wider range than the in-macroblock filter — H.264 did this, but only for intra edges.

  Third, VP8′s filter omits most of the adaptive strength mechanics inherent in H.264′s filter. Its only adaptation is that it skips filtering on p16x16 blocks with no coefficients. This may be responsible for the high blurriness of VP8′s loop filter : it will run over and over and over again on all parts of a macroblock even if they are unchanged between frames (as long as some other part of the macroblock is changed). H.264′s, by comparison, is strength-adaptive based on whether DCT coefficients exist on either side of a given edge and based on the motion vector delta and reference frame delta across said edge. Of course, skipping this strength calculation saves some decoding time as well.

  Update :
  05:28 < derf> Gumboot : You’ll be disappointed to know they got the loop filter ordering wrong again.
  05:29 < derf> Dark_Shikari : They ordered it such that you have to process each macroblock in full before processing the next one.

  Verdict on Loop Filter : Definitely worse compression-wise than H.264′s due to the lack of adaptive strength. Especially with the “fast” mode, might be significantly faster. I worry about it being too blurry.

  Overall verdict on the VP8 video format

  Overall, VP8 appears to be significantly weaker than H.264 compression-wise. The primary weaknesses mentioned above are the lack of proper adaptive quantization, lack of B-frames, lack of an 8×8 transform, and non-adaptive loop filter. With this in mind, I expect VP8 to be more comparable to VC-1 or H.264 Baseline Profile than with H.264. Of course, this is still significantly better than Theora, and in my tests it beats Dirac quite handily as well.

  Supposedly Google is open to improving the bitstream format — but this seems to conflict with the fact that they got so many different companies to announce VP8 support. The more software that supports a file format, the harder it is to change said format, so I’m dubious of any claim that we will be able to spend the next 6-12 months revising VP8. In short, it seems to have been released too early : it would have been better off to have an initial period during which revisions could be submitted and then a big announcement later when it’s completed.

  Update : it seems that Google is not open to changing the spec : it is apparently “final”, complete with all its flaws.

  In terms of decoding speed I’m not quite sure ; the current implementation appears to be about 16% slower than ffmpeg’s H.264 decoder (and thus probably about 25-35% slower than state-of-the-art decoders like CoreAVC). Of course, this doesn’t necessarily say too much about what a fully optimized implementation will reach, but the current one seems to be reasonably well-optimized and has SIMD assembly code for almost all major DSP functions, so I doubt it will get that much faster.

  I would expect, with equally optimized implementations, VP8 and H.264 to be relatively comparable in terms of decoding speed. This, of course, is not really a plus for VP8 : H.264 has a great deal of hardware support, while VP8 largely has to rely on software decoders, so being “just as fast” is in many ways not good enough. By comparison, Theora decodes almost 35% faster than H.264 using ffmpeg’s decoder.

  Finally, the problem of patents appears to be rearing its ugly head again. VP8 is simply way too similar to H.264 : a pithy, if slightly inaccurate, description of VP8 would be “H.264 Baseline Profile with a better entropy coder”. Even VC-1 differed more from H.264 than VP8 does, and even VC-1 didn’t manage to escape the clutches of software patents. It’s quite possible that VP8 has no patent issues, but until we get some hard evidence that VP8 is safe, I would be cautious. Since Google is not indemnifying users of VP8 from patent lawsuits, this is even more of a potential problem. Most importantly, Google has not released any justifications for why the various parts of VP8 do not violate patents, as Sun did with their OMS standard : such information would certainly cut down on speculation and make it more clear what their position actually is.

  But if luck is on Google’s side and VP8 does pass through the patent gauntlet unscathed, it will undoubtedly be a major upgrade as compared to Theora.

  Addendum A : On2′s VP8 Encoder and Decoder

  This post is primarily aimed at discussing issues relating to the VP8 video format. But from a practical perspective, while software can be rewritten and improved, to someone looking to use VP8 in the near future, the quality (both code-wise, compression-wise, and speed-wise) of the official VP8 encoder and decoder is more important than anything I’ve said above. Thus, after reading through most of the code, here’s my thoughts on the software.

  Initially I was intending to go easy on On2 here ; I assumed that this encoder was in fact new for VP8 and thus they wouldn’t necessarily have time to make the code high-quality and improve its algorithms. However, as I read through the encoder, it became clear that this was not at all true ; there were comments describing bugfixes dating as far back as early 2004. That’s right : this software is even older than x264 ! I’m guessing that the current VP8 software simply evolved from the original VP7 software. Anyways, this means that I’m not going to go easy on On2 ; they’ve had (at least) 6 years to work on VP8, and a much larger dev team than x264′s to boot.

  Before I tear the encoder apart, keep in mind that it isn’t bad. In fact, compression-wise, I don’t think they’re going to be able to get it that much better using standard methods. I would guess that the encoder, on slowest settings, is within 5-10% of the maximum PSNR that they’ll ever get out of it. There’s definitely a whole lot more to be had using unusual algorithms like MB-tree, not to mention the complete lack of psy optimizations — but at what it tries to do, it does pretty decently. This is in contrast to the VP3 encoder, which was a pile of garbage (just ask any Theora dev).

  Before I go into specific components, a general note on code quality. The code quality is much better than VP3, though there’s still tons of typos in the comments. They also appear to be using comments as a form of version control system, which is a bit bizarre. The assembly code is much worse, with staggering levels of copy-paste coding, some completely useless instructions that do nothing at all, unaligned loads/stores to what-should-be aligned data structures, and a few functions that are simply written in unfathomably roundabout (and slower) ways. While the C code isn’t half bad, the assembly is clearly written by retarded monkeys. But I’m being unfair : this is way better than with VP3.

  Motion estimation : Diamond, hex, and exhaustive (full) searches available. All are pretty naively implemented : hexagon, for example, performs a staggering amount of redundant work (almost half of the locations it searches are repeated !). Full is even worse in terms of inefficiency, but it’s useless for all but placebo-level speeds, so I’m not really going to complain about that.

  Subpixel motion estimation : Straightforward iterative diamond and square searches. Nothing particularly interesting here.

  Quantization : Primary quantization has two modes : a fast mode and a slightly slower mode. The former is just straightforward deadzone quant, while the latter has a bias based on zero-run length (not quite sure how much this helps, but I like the idea). After this they have “coefficient optimization” with two modes. One mode simply tries moving each nonzero coefficient towards zero ; the slow mode tries all 2^16 possible DCT coefficient rounding permutations. Whoever wrote this needs to learn what trellis quantization (the dynamic programming solution to the problem) is and stop using exponential-time algorithms in encoders.

  Ratecontrol (frame type handling) : Relies on “boosting” the quality of golden frames and “alt-ref” frames — a concept I find extraordinarily dubious because it means that the video will periodically “jump” to a higher quality level, which looks utterly terrible in practice. You can see the effect in this graph of PSNR ; every dozen frames or so, the quality “jumps”. This cannot possibly look good in motion.

  Ratecontrol (overall) : Relies on a purely reactive ratecontrol algorithm, which probably will not do very well in difficult situations such as hard-CBR and tight buffer constraints. Furthermore, it does no adaptation of the quantizer within the frame (e.g. in the case that the frame overshot the size limitations ratecontrol put on it). Instead, it relies on re-encoding the frame repeatedly to reach the target size — which in practice is simply not a usable option for two reasons. In low-latency situations where one can’t have a large delay, re-encoding repeatedly may send the encoder way behind time-wise. In any other situation, one can afford to use frame-based threading, a much faster algorithm for multithreaded encoding than the typical slice-based threading — which makes re-encoding impossible.

  Loop filter : The encoder attempts to optimize the loop filter parameters for maximum PSNR. I’m not quite sure how good an idea this is ; every example I’ve seen of this with H.264 ends up creating very bad (often blurry) visual results.

  Overall performance : Even on the absolute fastest settings with multithreading, their encoder is slow. On my 1.6Ghz Core i7 it gets barely 26fps encoding 1080p ; not even enough to reliably do real-time compression. x264, by comparison, gets 101fps at its fastest preset “ultrafast”. Now, sure, I don’t expect On2′s encoder to be anywhere near as fast as x264, but being unable to stream HD video on a modern quad-core system is simply not reasonable in 2010. Additionally, the speed options are extraordinarily confusing and counterintuitive and don’t always seem to work properly ; for example, fast encoding mode (–rt) seems to be ignored completely in 2-pass.

  Overall compression : As said before, compression-wise the encoder does a pretty good job with the spec that it’s given. The slower algorithms in the encoder are clearly horrifically unoptimized (see the comments on motion search and quantization in particular), but they still work.

  Decoder : Seems to be straightforward enough. Nothing jumped out at me as particularly bad, slow, or otherwise, besides the code quality issues mentioned above.

  Practical problems : The encoder and decoder share a staggering amount of code. This means that any bug in the common code will affect both, and thus won’t be spotted because it will affect them both in a matching fashion.  This is the inherent problem with any file format that doesn’t have independent implementations and is defined by a piece of software instead of a spec : there are always bugs. RV40 had a hilarious example of this, where a typo of “22″ instead of “33″ resulted in quarter-pixel motion compensation being broken. Accordingly, I am very dubious of any file format defined by software instead of a specification. Google should wait until independent implementations have been created before setting the spec in stone.

  Update : it seems that what I forsaw is already coming true :

  <derf> gmaxwell : It survives it with a patch that causes artifacts because their encoder doesn’t clamp MVs properly.
  <gmaxwell> ::cries: :
  <derf> So they reverted my decoder patch, instead of fixing the encoder.
  <gmaxwell> “but we have many files encoded with this !”
  <gmaxwell> so great.. single implementation and it depends on its own bugs.

  This is just like Internet Explorer 6 all over again — bugs in the software become part of the “spec” !

  Hard PSNR numbers :
  (Source/target bitrate are the same as in my upcoming comparison.)
  x264, slowest mode, High Profile : 29.76103db ( 28% better than VP8)
  VP8, slowest mode : 28.37708db ( 8.5% better than x264 baseline)
  x264, slowest mode, Baseline Profile : 27.95594db

  Note that these numbers are a “best-case” situation : we’re testing all three optimized for PSNR, which is what the current VP8 encoder specializes in as well. This is not too different from my expectations above as estimated from the spec itself ; it’s relatively close to x264′s Baseline Profile.

  Keep in mind that this is not representative of what you can get out of VP8 now, but rather what could be gotten out of VP8. PSNR is meaningless for real-world encoding — what matters is visual quality — so hopefully if problems like the adaptive quantization issue mentioned previously can be overcome, the VP8 encoder could be improved to have x264-level psy optimizations. However, as things stand…

  Visual results : Unfortunately, since the current VP8 encoder optimizes entirely for PSNR, the visual results are less than impressive. Here’s a sampling of how it compares with some other encoders. Source and bitrate are the same as above ; all encoders are optimized for optimal visual quality wherever possible. And apparently given some of the responses to this part, many people cannot actually read ; the bitrate is (as close as possible to) the same on all of these files.

  Update : I got completely slashdotted and my few hundred gigs of bandwidth ran out in mere hours. The images below have been rehosted, so if you’ve pasted the link somewhere else, check below for the new one.

  VP8 (On2 VP8 rc8) (source) (Note : I recently realized that the official encoder doesn’t output MKV, so despite the name, this file is actually a VP8 bitstream wrapped in IVF, as generated by ivfenc. Decode it with ivfdec.)
  H.264 (Recent x264) (source)
  H.264 Baseline Profile (Recent x264) (source)
  Theora (Recent ptalabvorm nightly) (source)
  Dirac (Schroedinger 1.0.9) (source)
  VC-1 (Microsoft VC-1 SDK) (source)
  MPEG-4 ASP (Xvid 1.2.2) (source)

  The quality generated by On2′s VP8 encoder will probably not improve significantly without serious psy optimizations.

  One further note about the encoder : currently it will drop frames by default, which is incredibly aggravating and may cause serious problems. I strongly suggest anyone using it to turn the frame-dropping feature off in the options.

  Addendum B : Google’s choice of container and audio format for HTML5

  Google has chosen Matroska for their container format. This isn’t particularly surprising : Matroska is one of the most widely used “modern” container formats and is in many ways best-suited to the task. MP4 (aka ISOmedia) is probably a better-designed format, but is not very flexible ; while in theory it can stick anything in a private stream, a standardization process is technically necessary to “officially” support any new video or audio formats. Patents are probably a non-issue ; the MP4 patent pool was recently disbanded, largely because nobody used any of the features that were patented.

  Another advantage of Matroska is that it can be used for streaming video : while it isn’t typically, the spec allows it. Note that I do not mean progressive download (a’la Youtube), but rather actual streaming, where the encoder is working in real-time. The only way to do this with MP4 is by sending “segments” of video, a very hacky approach in which one is effectively sending a bunch of small MP4 files in sequence. This approach is used by Microsoft’s Silverlight “Smooth Streaming”. Not only is this an ugly hack, but it’s unsuitable for low-latency video. This kind of hack is unnecessary for Matroska. One possible problem is that since almost nobody currently uses Matroska for live streaming purposes, very few existing Matroska implementations support what is necessary to play streamed Matroska files.

  I’m not quite sure why Google chose to rebrand Matroska ; “WebM” is a silly name and Matroska is already pretty well-recognized as a brand.

  The choice of Vorbis for audio is practically a no-brainer. Even ignoring the issue of patents, libvorbis is still the best general-purpose open source audio encoder. While AAC is generally better at very low bitrates, there aren’t any good open source AAC encoders : faac is worse than LAME and ffmpeg’s AAC encoder is even worse. Furthermore, faac is not free software ; it contains code from the non-free reference encoder. Combined with the patent issue, nobody expected Google to pick anything else.

  Addendum C : Summary for the lazy

  VP8, as a spec, should be a bit better than H.264 Baseline Profile and VC-1. It’s not even close to competitive with H.264 Main or High Profile. If Google is willing to revise the spec, this can probably be improved.

  VP8, as an encoder, is somewhere between Xvid and Microsoft’s VC-1 in terms of visual quality. This can definitely be improved a lot.

  VP8, as a decoder, decodes even slower than ffmpeg’s H.264. This probably can’t be improved that much ; VP8 as a whole is similar in complexity to H.264.

  With regard to patents, VP8 copies too much from H.264 for comfort, no matter whose word is behind the claim of being patent-free. This doesn’t mean that it’s sure to be covered by patents, but until Google can give us evidence as to why it isn’t, I would be cautious.

  VP8 is definitely better compression-wise than Theora and Dirac, so if its claim to being patent-free does stand up, it’s a big upgrade with regard to patent-free video formats.

  VP8 is not ready for prime-time ; the spec is a pile of copy-pasted C code and the encoder’s interface is lacking in features and buggy. They aren’t even ready to finalize the bitstream format, let alone switch the world over to VP8.

  With the lack of a real spec, the VP8 software basically is the spec–and with the spec being “final”, any bugs are now set in stone. Such bugs have already been found and Google has rejected fixes.

  Google made the right decision to pick Matroska and Vorbis for its HTML5 video proposal.

  29.76103