I just posted (to our prerelease site) LICEcap 1.25 beta 3, which includes support for using transparency for smaller images. I had some fun debugging this (including some very stupid mistakes on my part that took about an entire day to debug, oops).

I also had some fun writing logic to decide what to do when a pixel could be encoded as transparent, but also could be well-represented by an indexed color. Iniitially I had it only use the indexed value if the previous pixel was indexed, but it ended up being quite a bit better to do track the occurence of transparent pixels and pixels of that index, and use the one that is more common. There is probably a better algorithm to use here, but that saw some good gains. For comparison, I ran 1.24 and 1.25 beta 3 at the same time for a stupid demo video. The 1.24 version was 2.5MB, the 1.25 beta 3 version was 1.5MB. WIN. It ultimately is highly dependent on the content, though, so I might look at trying some other things out...

Comment...

Very often in computer code integers are divided by powers of two, such as 2, 4, 8, 16, 32, 64, and so on). These divisions are much faster than dividing by other numbers (since computers represent the underlying number in binary). In C/C++, there are two ways this type of division is typically expressed: normal division (x/256), or a shift (x<<8). These two methods produce the same results for non-negative values, and depending on the meaning of the code in question, (x/256) is often more readable, and thus I use it regularly.

If x is signed and is negative, division rounds towards 0, whereas the shift rounds towards negative infinity, but in situations where rounding of negative values is not important, I had generally assumed that modern compilers would generate similarly efficient code (reducing x/256 into a single x86 sar instruction).

It turns out, testing with a two modern compilers (and one slightly out of date compiler), this is not the case.

Here is the C code:

  void b(int r);
  void f(int v)
  {
    int i;
    for (i=0;i<v/256;i++) b(v > 0 ? v/65536 : 0);
  }
  void f2(int v)
  {
    int i;
    for (i=0;i<(v>>8);i++) b(v > 0 ? v >> 16 : 0);
  }

Ideally, a compiler should generate identical code for each of these functions. In the case of the loop counter, if v is less than 0, how it is rounded makes no difference. In the case of the parameter to b(), the code v >> 16 is only evaluated if v is known to be above 0.

Let's look at the output of some compilers (removing decoration and unrelated code). I've marked some code as bold to signify instructions that could be eliminated (with slight changes to the surrounding instructions):

• Apple LLVM version 5.1 (clang-503.0.40) (based on LLVM 3.4svn)
  Xcode 5.1.1 on OSX 10.9.4
  Target: x86_64-apple-darwin13.3.0
  Command line flags: -O2 -fomit-frame-pointer

  f():                              # using divides
  	cmpl	$256, %edi
  	jl	LBB0_3            # if v is less than 256, skip loop
          # v is known to be 256 or greater. 
  	movl	%edi, %ebx
  	sarl	$31, %ebx         # ebx=0xffffffff if negative, 0 if non-negative
  	movl	%ebx, %r14d    
  	shrl	$24, %r14d        # r14d=0xff if negative, 0 if non-negative
  	addl	%edi, %r14d       # r14d = (v+255) if v negative, v if v non-negative
  	sarl	$8, %r14d         # r14d = v/256
  	shrl	$16, %ebx         # this will make ebx 65535 if v negative, 0 if v non-negative
  	addl	%edi, %ebx        # ebx = (v+65535) if v negative, v if v non-negative
  	sarl	$16, %ebx         # ebx = v/65536
          # interestingly, the optimizer used its knowledge of v being greater than 0 to remove the ternary conditional expression completely.
  	xorl	%ebp, %ebp
  LBB0_2:
  	movl	%ebx, %edi
  	callq	_b
  	incl	%ebp
  	cmpl	%r14d, %ebp
  	jl	LBB0_2
  LBB0_3:
  f2():                             # using shifts
  	movl	%edi, %ebp
  	sarl	$8, %ebp          # ebp = v>>8
  	testl	%ebp, %ebp
  	jle	LBB1_3            # if less than or equal to 0, skip
  	movl	%edi, %eax
  	sarl	$16, %eax         # eax = v>>16
  	xorl	%ebx, %ebx
  	testl	%edi, %edi
  	cmovgl	%eax, %ebx        # if v is greater than 0, set ebx to eax
          # the optimizer could also have removed the xorl/testl/cmovgl sequence as well
  LBB1_2:
  	movl	%ebx, %edi
  	callq	_b
  	decl	%ebp
  	jne	LBB1_2
  LBB1_3:
  

  In the first function (division), the LLVM optimizer appears to have removed the ternary expression (checking to see if v was greater than 0), likely because it knew that if the loop was running, v was greater than 0. Unfortunately, it didn't apply this knowledge to the integer divisions of v, which would have allowed it to not generate (substantial) rounding code.
  
  In the second function (shifts), LLVM wasn't required to generate rounding code (as C's >> maps to x86 sar directly), but it also didn't use the knowledge that v would be greater than 0.

• Microsoft (R) C/C++ Optimizing Compiler Version 18.00.30501 for x64
  Visual Studio Express 2013 for Windows Desktop on Windows 7
  Command line flags: /O2 (/Os produced different results, but nothing related to rounding)

  f():                              # using divides
          mov     eax, ecx
          mov     ebx, ecx
          cdq                       # set edx to 0xffffffff if v negative, 0 otherwise
          movzx   edx, dl           # set edx to 0xff if v negative, 0 otherwise
          add     eax, edx          # eax = v+255 if v negative, v otherwise
          sar     eax, 8            # eax = v/256
          test    eax, eax
          jle     SHORT $LN1@f      # skip loop if v/256 is less than or equal to 0
          mov     QWORD PTR [rsp+48], rdi
          mov     edi, eax          # edi is loop counter
  $LL3@f:
          test    ebx, ebx
          jle     SHORT $LN6@f      # if v is less than or equal to 0, jump to set eax to 0
          mov     eax, ebx
          cdq                       # set edx to 0xffffffff if v negative, 0 otherwise
          movzx   edx, dx           # set edx to 0xffff if v negative, 0 otherwise
          add     eax, edx          # eax = v+65535 if v negative, v otherwise
          sar     eax, 16           # eax = v/65536
          jmp     SHORT $LN7@f
  $LN6@f:
          xor     eax, eax
  $LN7@f:
          mov     ecx, eax
          call    b
          dec     rdi
          jne     SHORT $LL3@f
          mov     rdi, QWORD PTR [rsp+48]
  $LN1@f:
  f2():                             # using shifts
          mov     eax, ecx
          mov     ebx, ecx
          sar     eax, 8            # eax = v>>8
          test    eax, eax
          jle     SHORT $LN1@f2     # skip loop if v>>8 is less than or equal to 0
          mov     QWORD PTR [rsp+48], rdi
          mov     edi, eax
  $LL3@f2:
          test    ebx, ebx
          jle     SHORT $LN6@f2     # if v is less than or equal to 0, jump to set ecx to 0
          mov     ecx, ebx
          sar     ecx, 16           # ecx = v>>16
          jmp     SHORT $LN7@f2
  $LN6@f2:
          xor     ecx, ecx
  $LN7@f2:
          call    b
          dec     rdi
          jne     SHORT $LL3@f2
          mov     rdi, QWORD PTR [rsp+48]
  $LN1@f2:
  

  VS 2013 generates different rounding code for the division, using cdq/movzx (or cdq/and if shifting by something other than 8 or 16 bits).
  
  Also worth noting is that VS 2013 doesn't even bother moving the invariant ternary operator and (v/65536) or (v>>16) out of the loop. Ideally, it could move that calculation out of the loop, or remove the ternary operator completely. Ouch. I have to say, VS 2013 does seem to produce pretty good code overall, but I guess most of ours is heavily in floating point these days.

• gcc 4.4.5
  Linux x86_64
  Command line flags: -O2 -fomit-frame-pointer

  f():                              # using divides
          testl   %edi, %edi
          movl    %edi, %ebp        # ebp = edi = v
          leal    255(%rbp), %r12d  # r12d = v+255
          cmovns  %edi, %r12d       # set r12d to v, if v is non-negative (otherwise r12d was v+255)
          sarl    $8, %r12d         # r12d = v/256
          testl   %r12d, %r12d
          jle     .L14              # if r12d is less than or equal to 0, skip
          movl    %edi, %r14d
          xorl    %ebx, %ebx        # ebx is loop counter
          xorl    %r13d, %r13d
          sarl    $16, %r14d        # r14d = v>>16
  .L13:
          testl   %ebp, %ebp
          movl    %r13d, %edi
          cmovg   %r14d, %edi       # if v is greater than 0, use v>>16 instead of 0
          addl    $1, %ebx
          call    b
          cmpl    %r12d, %ebx
          jl      .L13
  .L14:
  f2():                             # using shifts
          movl    %edi, %r12d
          sarl    $8, %r12d
          testl   %r12d, %r12d
          movl    %edi, %ebp
          jle     .L6               # skip loop if (v>>8) is less than or equal to 0
          movl    %edi, %r14d
          xorl    %ebx, %ebx        # ebx is loop counter
          xorl    %r13d, %r13d
          sarl    $16, %r14d        # r14d = (v>>16)
  .L5:
          testl   %ebp, %ebp
          movl    %r13d, %edi
          cmovg   %r14d, %edi       # if v is greater than 0, use v>>16 instead of 0
          addl    $1, %ebx
          call    b
          cmpl    %r12d, %ebx
          jl      .L5
  .L6:
  

  gcc 4.4 does an interesting job, using lea to generate v+255, and then cmovns to replace it with v if v is non-negative. It doesn't bother generating rounding code for v/65536, but it does still generate rounding code for v/256, even though any non-positive result for v/256 is treated the same way throughout. Also, gcc doesn't eliminate the non-varying ternary expression, nor put the constant v/65536 or v>>16 outside of the loop.

I'm not sure what to say here -- modern compilers can generate a lot of really good code, especially looking at floating point and SSE, but this makes me feel as though some of the basics have been neglected. If I were a better programmer I'd go dig into LLVM and GCC and submit patches.

I should have also tested ICC, but I've spent enough time on this, and the only ICC version we use is old enough that I would just regret not using the latest.

For comparison, here is what I would like to see LLVM generate for f():

f():                              # using divides
	cmpl	$256, %edi
	jl	LBB0_3            # if v is less than 256, skip loop
	movl	%edi, %ebx
	movl	%edi, %ebp
	shrl	$8, %ebx          # ebx = v/256, since v is non-negative
	shrl	$16, %ebp         # ebp = v/65536, since v is non-negative
LBB0_2:
	movl	%ebp, %edi
	callq	_b
	decl    %ebx
	jnz	LBB0_2
LBB0_3:

2 Comments

I've been investigating, once again, the performance of drawing code-rendered RGBA bitmaps to NSViews in OSX. I found that on my Retina Macbook Pro (when the application was not in low-resolution legacy mode), calling CGContextSetInterpolationQuality with kCGInterpolationNone would cause CGContextDrawImage() to be more than twice as fast (with less filtering of the image, which was a fair tradeoff and often desired).

The above performance gain aside, I am still not satisfied with the bitmap drawing performance on recent OSX versions, which has led me to benchmark SWELL's blitting code. My test uses the LICE test application, with a screen full of lines, an opaque NSView, and 720x500 resolution.

OSX 10.6 vs 10.8 on a C2D iMac

My (C2D 2.93GHz) iMac running 10.6 easily runs the benchmark at close to 60 FPS, using about 45% of one core, with the BitBlt() call typically taking 1ms for each frame.

Here is a profile -- note that CGContextDrawImage() accounts for a modest 3.9% of the total CPU use:

It might be possible to reduce the work required by changing our bitmap representation from ABGR to RGBA (avoiding sseCGSConvertXXXX8888TransposeMask and performing a memcpy() instead), but in my opinion 1ms for a good sized blit (and less than 4% of total CPU time for this demo) is totally acceptable.

I then rebooted the C2D iMac into OSX 10.8 (Mountain Lion) for a similar test.

Running the same benchmark on the same hardware in Mountain Lion, we see that each call to BitBlt() takes over 6ms, the application struggles to exceed 57 FPS, and the CPU usage is much higher, at about 73% of a core.

Here is the time sampling of the CGContextDrawImage() -- in this case it accounts for 36% of the total CPU use!

Looking at the difference between these functions, it is obvious where most of the additional processing takes place -- within img_colormatch_read and CGColorTransformConvertData, where it apparently applies color matching transformations.

I'm happy that Apple cares about color matching, but to force it on (without allowing developers control over it) is wasteful. I'd much rather have the ability transform the colors before rendering, and be able to quickly blit to screen, than to have to have every single pixel pushed to the screen color transformed. There may be some magical way to pass the right colorspace value to CGCreateImage() to bypass this, but I have not found it yet (and I have spent a great deal of time looking, and trying things like querying the monitor's colorspace).

That's what OpenGL is for!
But wait, you say -- the preferred way to quickly draw to screen is OpenGL.

Updating a complex project to use OpenGL would be a lot of work, but for this test project I did implement a very naive OpenGL blit, which enabled an OpenGL context for the view and created a texture for drawing each frame, more or less like:

    glDisable(GL_TEXTURE_2D);
    glEnable(GL_TEXTURE_RECTANGLE_EXT);
    GLuint texid=0;
    glGenTextures(1, &texid);
    glBindTexture(GL_TEXTURE_RECTANGLE_EXT, texid);
    glPixelStorei(GL_UNPACK_ROW_LENGTH, sw);
    glTexParameteri(GL_TEXTURE_RECTANGLE_EXT, GL_TEXTURE_MIN_FILTER,  GL_LINEAR);
    glTexImage2D(GL_TEXTURE_RECTANGLE_EXT,0,GL_RGBA8,w,h,0,GL_BGRA,GL_UNSIGNED_INT_8_8_8_8, p);
    glBegin(GL_QUADS);
    glTexCoord2f(0.0f, 0.0f);
    glVertex2f(-1,1);
    glTexCoord2f(0.0f, h);
    glVertex2f(-1,-1);
    glTexCoord2f(w,h);
    glVertex2f(1,-1);
    glTexCoord2f(w, 0.0f);
    glVertex2f(1,1);
    glEnd();
    glDeleteTextures(1,&texid);
    glFlush();

The memory use when using OpenGL blitting increased by about 10MB, which may not sound like much, but if you are drawing to many views, the RAM use would potentially increase with each view.

I also tested the OpenGL implementation on 10.6, but it was significantly slower than CoreGraphics: 3ms per frame, nearly 60 FPS but CPU use was 60% of a core, so if you do ever implement OpenGL blitting, you will probably want to disable it for 10.6 and earlier.

Core 2 Duo?! That's ancient, get a new computer!
After testing on the C2D, I moved back to my modern quad-core i7 Retina Macbook Pro running 10.9 (Mavericks) and did some similar tests.

• Normal: 12-14ms per frame, 46 FPS, 70% of a core CPU use
• Normal, in "Low Resolution" mode: 6-7ms per frame, 58FPS, 60% of a core CPU use
• Normal, without the kCGInterpolationNone: 29ms per frame, 29 FPS, 70% of a core CPU use
• Normal, in "Low Resolution" mode, without kCGInterpolationNone: same as with kCGInterpolationNone.
• GL: 1-2ms per frame, 57 FPS, 37% of a core CPU
• GL, in "Low Resolution" mode: 1-2ms per frame, 57 FPS, 40% of a core CPU

Let's see where the time is spent in the "Normal, Low Resolution" mode:

This looks very similar to the 10.8, non-retina rendering, though some function names have changed. There is the familiar img_colormatch_read/CGColorTransformConvertData call which is eating a good chunk of CPU. The ripc_RenderImage/ripd_Mark/argb32_image stack is similar to 10.8, and reasonable in CPU cycles consumed.

Looking at the Low Resolution mode, it really does behave similar to that of 10.8 (though it's depressing to see that it still takes as long to run on an i7 as 10.8 did on a C2D, hmm). Let's look at the full-resolution Retina mode:

img_colormatch_read is present once again, but what's new is that ripc_RenderImage/ripd_Mark/argb32_image have a new implementation, calling argb32_image_mark_RGB24 -- and argb32_image_mark_RGB24 is a beast! It uses more CPU than just about anything else. What is going on there?

Conclusions
If you ever feel as if modern OSX versions have gotten slower when it comes to updating the screen, you would be right. The basic method of drawing ixels rendered in a platform-independent fashion to screen has gotten significantly slower since Snow Leopard, most likely in the name of color-accuracy. In my opinion this is an oversight on Apple's part, and they should extend the CoreGraphics APIs to allow manual application of color correction.

Additionally, I'm suspicious that something odd is going on within the function argb32_image_mark_RGB24, which appears to only be used on Retina displays, and that the performance of that function should be evaluated. Improving the efficiency of that function would have a positive impact on the performance of many third party applications (including REAPER).

If anybody has an interest in duplicating these results or doing further testing, I have pushed the updates to the LICE test application to our WDL git repository (see WDL/lice/test/).

Update: July 3, 2014
After some more work, I've managed to get the CPU use down to a respectable level in non-Retina mode (10.8 on the iMac, 10.9/Low Resolution on the Retina MBP), by using the system monitor's colorspace:

    CMProfileRef systemMonitorProfile = NULL;
    CMError getProfileErr = CMGetSystemProfile(&systemMonitorProfile);
    if(noErr == getProfileErr)
    {
      cs = CGColorSpaceCreateWithPlatformColorSpace(systemMonitorProfile);
      CMCloseProfile(systemMonitorProfile);
    }

However, this mode is appears to be slower on the Retina MBP in high resolution mode, as it calls argb32_image_mark_RGB32 instead of argb32_image_mark_RGB24 (presumably operating on my buffer directly rather than the intermediate colorspace-converted buffer), which is even slower.

Update: July 3, 2014, later
OK, if you provide a bitmap that is twice the size of the drawing rect, you can avoid argb32_image_mark_RGBXX, and get the Retina display to update in about 5-7ms, which is a good improvement (but by no means impressive, given how powerful this machine is). I made a very simple software scaler (that turns each pixel into 4), and it uses very little CPU. So this is acceptable as a workaround (though Apple should really optimize their implementation). We're at least around 6ms, which is way better than 12-14ms (or 29ms which is where we were last week!), but there's no reason this can't be faster. As a nice side effect, I'm adding SWELL_IsRetinaDC(), so we can start making some things Retina aware -- JSFX GUIs would be a good place to start...

1 Comment

Comment...

Comment...

2 Comments

midi2osc, as mentioned in the last post, got some updates which made its name obsolete, particularly the ability to send MIDI and receive OSC, so it has now been renamed OSCII-bot. Other mildly interesting updates:

• OSX version
• Can load multiple scripts, which run independently but can share hardware
• Better string syntax (normal quotes rather than silly {} etc), user strings identified by values 0..1023
• Better string manipulation APIs (sprintf(), strcpy(), match(), etc).
• match() and oscmatch(), which can be used for simple regular expressions with the ability to extract values
• Ability to detect stale devices and reopen them
• Scripts can output text to the newly resizeable console, including basic terminal emualtion (\r, and \33[2J for clear)
• Vastly improved icon

Making installers and web pages is too much of a pain for a small project, so:

Here's another project, this one took less than 24 hours to make. It's a little win32 program called "midi2osc" (license: GPL, binary included, code requires WDL to compile), and it simply listens to any number of MIDI devices, and broadcasts to any number of destinations via OSC-over-UDP.

MIDI and OSC use completely different types of encoding -- MIDI consists of 1-3 byte sequences (excluding sysex), and OSC is encoded as a string and any number of values (strings, floats, integers, whatever). It would be very easy to make a simplistic conversion of every MIDI event, such as 90 53 7f being converted to "/note/on/53" with an integer value of 7f, and so on. This would be useful, but also might be somewhat limited.

In order to make this as useful as possible, I made it use EEL2 to enable user scripting of events. EEL2 is a fast scripting engine designed for floating point computation, that was originally developed as part of Nullsoft AVS, and evolved as part of our Jesusonic/JSFX code. EEL2 compiles extremely quickly to native code, and can have context that is used by code running in multiple threads simultaneously.

For this project the EEL2 syntax was extended slightly, via the use of a preprocessor, so that you can specify format strings for OSC. For example, you can tell REAPER to set a track's volume via:

    oscfmt0 = trackindex;
    oscsend(destination, { "/track/%.0f/volume" }, 0.5);

• If the MIDI input device of the R24 is not opened, audio playback is not reliable. This includes when listening to Winamp or watching YouTube. So basically, for this thing to be useful, I need something to keep hitting the MIDI device constantly. So for you Zoom R24 win64 users who have this problem, midi2osc might be able to fix your problems.
• If REAPER crashes or is otherwise debugged with the MIDI device open, the process hangs and it's a pain to continue. Moving the MIDI control to a separate process that can run in the system tray = win.
• (not midi2osc related): I wish the drum pads would send MIDI too... *ahem*

// @input lines:
// usage: @input devicenameforcode "substring match" [skip]
// can use any number of inputs. devicenameforcode must be unique, if you specify multiple @input lines
// with common devicenameforcode, it will use the first successful line and ignore subsequent lines with that name
// you can use any number of devices, too
@input r24 "ZOOM R"
// @output lines
// usage: @output devicenameforcode "127.0.0.1:8000" [maxpacketsize] [sleepamt]
// maxpacketsize is 1024 by default, can lower or raise depending on network considerations
// sleepamt is 10 by default, sleeps for this many milliseconds after each packet. can be 0 for no sleep.
@output localhost "127.0.0.1:8000"
@init
// called at init-time
destdevice = localhost; // can also be -1 for broadcast
// 0= simplistic /track/x/volume, /master/volume
// 1= /r24/rawfaderXX (00-09)
// 2= /action/XY/cc/soft (tracks 1-8), master goes to /r24/rawfader09
fader_mode=2;
@timer
// called around 100Hz, after each block of @msg
@msg
// special variables:
// time (seconds)
// msg1, msg2, msg3 (midi message bytes)
// msgdev == r24  // can check which device, if we care
(msg1&0xf0) == 0xe0 ? (
  // using this to learn for monitoring fx, rather than master track
  fader_mode > 0 ? (
     fmtstr = { f/r24/rawfader%02.0f }; // raw fader
     oscfmt0 = (msg1&0xf)+1;
     fader_mode > 1 && oscfmt0 != 9 ? (
       fmtstr = { f/action/%.0f/cc/soft }; // this is soft-takeover, track 01-08 volume
       oscfmt0 = ((oscfmt0-1) * 8) + 20;
     );
     val=(msg2 + (msg3*128))/16383;
     val=val^0.75;
     oscsend(destdevice,fmtstr,val);
  ) : (
     fmtstr = (msg1&0xf) == 8 ? { f/master/volume } : { "f/track/%.0f/volume"};
     oscfmt0 = (msg1&0xf)+1;
     oscsend(destdevice,fmtstr,(msg2 + (msg3*128))/16383);
  );
);
msg1 == 0x90 ? (
  msg2 == 0x5b ? oscsend(destdevice, { b/rewind }, msg3>64);
  msg2 == 0x5c ? oscsend(destdevice, { b/forward }, msg3>64);
  msg3>64 ? (
    oscfmt0 = (msg2&7) + 1;
    msg2 < 8 ?  oscsend(destdevice, { t/track/%.0f/recarm/toggle }, 0) :
      msg2 < 16 ?  oscsend(destdevice, { t/track/%.0f/solo/toggle }, 0) :
        msg2 < 24 ?  oscsend(destdevice, { t/track/%.0f/mute/toggle }, 0) : 
    (
      msg2 == 0x5e ? oscsend(destdevice, { b/play }, 1);
      msg2 == 0x5d ? oscsend(destdevice, { b/stop }, 1);
      msg2 == 0x5f ? oscsend(destdevice, { b/record }, 1);
    )
  );
);
msg1 == 0xb0 ? (
  msg2 == 0x3c ? (
    oscsend(destdevice, { f/action/992/cc/relative }, ((msg3&0x40) ? -1 : 1));
  );
);

Today I wrote a proof of concept program called "mikmod2rpp" (requires libmikmod and WDL). Most of the code in there is based on the Gravis Ultrasound driver for MikMod, but instead of uploading the sample data to a soundcard, it writes it to .wav files, and instead of playing the samples on the soundcard, it writes a REAPER project file (RPP). It is not perfect, does many things wrong, but it's somewhat usable.

Here are a couple of test projects (which include the mod/xm as well as the converted files): spathi.mod (from SC2), and resistance_is_futile.xm.

In case there's any question, it turns out the tracker format is incredibly efficient for both editing and playback... compared to a DAW.

1 Comment

Digging the album "Recreation" by Papier Tigre. Here's a pretty good live video (bad mix from one side of the stage, but, it's still quite listenable IMO):

Comment...

Comment...