Recherche avancée

Sur d’autres sites (5834)

• Bit-field badness

  30 janvier 2010, par Mans — Compilers, Optimisation

  Consider the following C code which is based on an real-world situation.

  struct bf1_31 
      unsigned a:1 ;
      unsigned b:31 ;
   ;
  void func(struct bf1_31 *p, int n, int a)
  
  
  
      int i = 0 ;
  
      do 
  
          if (p[i].a)
  
              p[i].b += a ;
  
       while (++i < n) ;
  
  
  

  How would we best write this in ARM assembler ? This is how I would do it :
  

  func :
          ldr     r3,  [r0], #4
          tst     r3,  #1
          add     r3,  r3,  r2,  lsl #1
          strne   r3,  [r0, #-4]
          subs    r1,  r1,  #1
          bgt     func
          bx      lr
  

  The add instruction is unconditional to avoid a dependency on the comparison. Unrolling the loop would mask the latency of the ldr instruction as well, but that is outside the scope of this experiment.

  Now compile this code with gcc -march=armv5te -O3 and watch in horror :

  func :
          push    r4
          mov     ip, #0
          mov     r4, r2
  loop :
          ldrb    r3, [r0]
          add     ip, ip, #1
          tst     r3, #1
          ldrne   r3, [r0]
          andne   r2, r3, #1
          addne   r3, r4, r3, lsr #1
          orrne   r2, r2, r3, lsl #1
          strne   r2, [r0]
          cmp     ip, r1
          add     r0, r0, #4
          blt     loop
          pop     r4
          bx      lr
  

  This is nothing short of awful :

  • The same value is loaded from memory twice.
  • A complicated mask/shift/or operation is used where a simple shifted add would suffice.
  • Write-back addressing is not used.
  • The loop control counts up and compares instead of counting down.
  • Useless mov in the prologue ; swapping the roles or r2 and r4 would avoid this.
  • Using lr in place of r4 would allow the return to be done with pop {pc}, saving one instruction (ignoring for the moment that no callee-saved registers are needed at all).

  Even for this trivial function the gcc-generated code is more than twice the optimal size and slower by approximately the same factor.

  The main issue I wanted to illustrate is the poor handling of bit-fields by gcc. When accessing bitfields from memory, gcc issues a separate load for each field even when they are contained in the same aligned memory word. Although each load after the first will most likely hit L1 cache, this is still bad for several reasons :

  • Loads have typically two or three cycles result latency compared to one cycle for data processing instructions. Any bit-field can be extracted from a register with two shifts, and on ARM the second of these can generally be achieved using a shifted second operand to a following instruction. The ARMv6T2 instruction set also adds the SBFX and UBFX instructions for extracting any signed or unsigned bit-field in one cycle.
  • Most CPUs have more data processing units than load/store units. It is thus more likely for an ALU instruction than a load/store to issue without delay on a superscalar processor.
  • Redundant memory accesses can trigger early flushing of store buffers rendering these less efficient.

  No gcc bashing is complete without a comparison with another compiler, so without further ado, here is the ARM RVCT output (armcc --cpu 5te -O3) :

  func :
          mov     r3, #0
          push    r4, lr
  loop :
          ldr     ip, [r0, r3, lsl #2]
          tst     ip, #1
          addne   ip, ip, r2, lsl #1
          strne   ip, [r0, r3, lsl #2]
          add     r3, r3, #1
          cmp     r3, r1
          blt     loop
          pop     r4, pc
  

  This is much better, the core loop using only one instruction more than my version. The loop control is counting up, but at least this register is reused as offset for the memory accesses. More remarkable is the push/pop of two registers that are never used. I had not expected to see this from RVCT.

  Even the best compilers are still no match for a human.

• Grabbing rtsp/rtp stream with libavformat

  19 janvier 2012, par T0ha

  world.
  I'm trying to grab rtsp mjpeg stream from IP-camera (realtime) as described in http://www.inb.uni-luebeck.de/ boehme/using_libavcodec.html, but ported to newer version.
  It works well with mpeg file (loading it full as one AVPacket), but working with stream, avcodec_decode_video2 returns -1 (error). AVPacket in this case is a part of a frame.

  How can I fix this ?

• 2011 In Open Source Multimedia

  5 janvier 2012, par Multimedia Mike — Open Source Multimedia

  Sometimes I think that the pace of multimedia technology is slowing down. Obviously, I’m not paying close enough attention. I thought I would do a little 2011 year-end review of what happened in the world of open source multimedia, mainly for my own benefit. Let me know in the comments what I missed.

  The Split
  The biggest deal in open source multimedia was the matter of the project split. Where once stood one project (FFmpeg) there now stands two (also Libav). Where do things stand with the projects now ? Still very separate but similar. Both projects obsessively monitor each other’s git commits and prodigiously poach each other’s work, both projects being LGPL and all. Most features that land in one code base end up in the other. Thus, I refer to FFmpeg and Libav collectively as “the projects”.

  Some philosophical reasons for the split included project stagnation and development process friction. Curiously, these problems are fond memories now and the spirit of competition has pushed development forward at a blinding pace.

  People inside the project have strong opinions about the split ; that’s understandable. People outside the project have strong opinions about the split ; that’s somewhat less understandable, but whatever. After 5 years of working for Adobe on the Flash Player (a.k.a. the most hated software in all existence if internet nerds are to be believed on the matter), I’m so over internet nerd drama.

  For my part, I just try to maintain some appearance of neutrality since I manage some shared resources for the open source multimedia community (like the wiki and samples repo) and am trying to keep them from fracturing as well.

  Apple and Open Source
  It was big news that Apple magnanimously open sourced their lossless audio codec. That sets a great example and precedent.

  New Features
  I mined the 'git log' of the projects in order to pick out some features that were added during 2011.

  First off, Apple’s ProRes video codec was reverse engineered and incorporated into the multimedia libraries. And for some weird reason, this is an item that made the rounds in the geek press. I’m not entirely sure why, but it may have something to do with inter-project conflict. Anyway, here is the decoder in action, playing a video of some wild swine, one of the few samples we have :

  
  
  

  Other new video codecs included a reverse engineered Indeo 4 decoder. Gotta catch ‘em all ! That completes our collection of Indeo codecs. But that wasn’t enough– this year, we got a completely revised Indeo 3 decoder (the previous one, while functional, exhibited a lot of code artifacts betraying a direct ASM ->C translation). Oh, and many thanks to Kostya for this gem :

  
  
  

  That’s the new Origin Xan decoder (best known for Wing Commander IV cinematics) in action, something I first started reverse engineering back in 2002. Thanks to Kostya for picking up my slack yet again.

  Continuing with the codec section, there is a decoder for Adobe Flash Screen Video 2 — big congrats on this ! One of my jobs at Adobe was documenting this format to the outside world and I was afraid I could never quite make it clear enough to build a complete re-implementation. But the team came through.

  Let’s see, there are decoders for VBLE video, Ut Video, Windows Media Image (WMVP/WMP2), Bink audio version ‘b’, H.264 4:2:2 intra frames, and MxPEG video. There is a DPX image encoder, a Cirrus Logic AccuPak video encoder, and a v410 codec.

  How about some more game stuff ? The projects saw — at long last — an SMJPEG demuxer. This will finally allow usage and testing of the SMJPEG IMA ADPCM audio decoder I added about a decade ago. Funny story behind that– I was porting all of my decoders from xine which included the SMJPEG ADPCM. I just never quite got around to writing a corresponding demuxer. Thanks to Paul Mahol for taking care of that.

  Here’s a DFA playback system for a 1995 DOS CD-ROM title called Chronomaster. No format is too obscure, nor its encoded contents too cheesy :

  
  
  

  There’s now a demuxer for a format called XMV that was (is ?) prevalent on Xbox titles. Now the projects can handle FMV files from many Xbox games, such as Thrillville.

  
  
  

  The projects also gained the ability to play BMV files. I think this surfing wizard comes from Discworld II. It’s non-computer-generated animation at a strange resolution.

  
  
  

  More demuxers : xWMA, PlayStation Portable PMP format, and CRI ADX format ; muxer for OpenMG audio and LATM muxer/demuxer.

  One more thing : an AVX-optimized fast Fourier transform (FFT). If you have a machine that supports AVX, there’s no way you’ll even notice the speed increase of a few measly FFT calls for audio coding/decoding, but that’s hardly the point. The projects always use everything on offer for any CPU.

  Please make me aware of features that I missed in the list !

  Continuous Testing
  As a result of the split, each project has its own FATE server, one for FFmpeg and one for Libav. As of the new year, FFmpeg has just over 1000 tests while Libav had 965. This is one area where I’m obviously ecstatic to see competition. Some ad-hoc measurements on my part indicate that the total code coverage via the FATEs has not appreciably increased. But that’s a total percentage. Both the test count and the code count have been steadily rising.

  Google Summer of Code and Google Code-In
  Once again, the projects were allowed to participate in the Google Summer of Code as well as Google Code-In. I confess that I didn’t keep up with these too carefully (and Code-In is still in progress as of this writing). I do know that the project split occurred after FFmpeg had already been accepted for GSoC season 2011 and the admins were gracious enough to allow FFmpeg and Libav to allow both projects to participate in the same slot as long as they could both be mature about it.

  Happy New Year
  Let’s see what we can accomplish in 2012.