AviSynth 2.6.0 Alpha3 [May 25th, 2011] - Page 17

IanB · 8th July 2012, 11:36

@forclip,

With the script as quoted, and substituting a 384Kbit stereo .AC3 file I had lying around, I see no problem with any versions of Avisynth.dll from 2.5.8.5, 2.6.0.0, 2.6.0.1, 2.6.0.2, current CVS and my current development version. I used a fairly old modified NicAudio.dll I built myself based on version 2.0.0.5. circa 2007.

I tried a single pass linear save of the output stream as well as random spot playback.

From your avs_audio.png image, the bass25.wav sample has a big jump in level at around 2:10, this might be bugs with decoding your file with seeking. Add the EnsureVBRMP3Sync() filter to your NicAC3Source() and observe any change in behaviour. Also try to reproduce the results using ColorBars() in place of NicAC3Source() for the audio source.

forclip · 9th July 2012, 01:28

Ok. With ColorBars() I also can reproduce this strange slowness issue, however audio output seems to be fine and it is not as slow as with the real sources. Using this script:

Code:

ColorBars().Trim(0, 50000)
MergeChannels(last, last, last)
MergeChannels(last, last)
ConvertAudioTo16bit()

with AviSynth 2.6 from 2009 it takes ~18sek to find the peaks. With more recent version(s) it takes ~23sek. Also, to get broken output my previous script can be simplified like this:

Code:

video = BlankClip(length=4248, fps=25.000, color=$000000)
audio = bassAudioSource("E:\Temp\1804_0.ac3")#.EnsureVBRMP3Sync()
AudioDub(video, audio)

MergeChannels(last, last)
ConvertAudioTo16bit()

The result is attached (and I will PM you with the source file). Uncommenting EnsureVBRMP3Sync() makes the things SLOW AS HELL, it takes something aboud 15 minutes to process! However, the output seems to be OK, but I've tried it only once since it is really VERY-VERY slow. But as I mentioned in my first post, this slowness sometimes appears, sometimes not, but most of the time it appears. Replacing avisynth.dll to version from 2009 - and exactly the same script ALWAYS takes only 4-5sek to precess, w\o any slowness. Before switching to AviSynth 2.6 (from 2.5.8 MT) I've never heard from my users about any problems related to audio processing (except "audio crash" that in 2.6 is fixed), now after we are using AviSynt 2.6, some of them reporting me that audio processing is SLOW from now. I can reproduce it on my side, also I can reproduce it in MeGUI, so this is not my app fault. This strange slowness can be reproduced with avs2wav too, but output seems to be OK, or I've tested it not very well..

Any ideas?

tebasuna51 · 9th July 2012, 09:12

Please, send me also the source file 1804_0.ac3 to see if I can detect any trouble decoding whit NicAc3Source/bassAudioSource.

I don't know for what EnsureVBRMP3Sync() can make differences decoding CBR audio, like ac3 must be.

IanB · 10th July 2012, 05:02

@forclip,

Between 2.6.0.1 and 2.6.0.2 the ac_expected_next update line was moved later in the code, so for the no audio cache state the skipping chunks case was constantly being incorrectly triggered. This led to all the default audio caches needlessly becoming enabled resulting in lots of extra memcpy'ing happening, slowing things down. Not sure if it's the whole cause but it does have quite a performance impact.

Code:

...
  long _cs = ac_currentscore;
  if (start < ac_expected_next)
    _cs = InterlockedExchangeAdd(&ac_currentscore, -5) - 5;  // revisiting old ground - a cache could help
  else if (start > ac_expected_next)
    _cs = InterlockedDecrement(&ac_currentscore);            // skiping chunks - a cache might not help
  else // (start == ac_expected_next)
    _cs = InterlockedIncrement(&ac_currentscore);            // strict linear reading - why bother with a cache

...

>>  ac_expected_next = start + count;

  if (h_audiopolicy == CACHE_AUDIO_NONE || h_audiopolicy == CACHE_AUDIO_NOTHING) {
    child->GetAudio(buf, start, count, env);
    return;  // We are ok to return now!
  }

...

<<  ac_expected_next = start + count;

  while (count>maxsamplecount) {    //is cache big enough?

@tebasuna51,

EnsureVBRMP3Sync() adds a 1MB audio cache and causes a high penalty for any out of order accesses outside the audio cache (a seek to zero plus a linear read up to the new position). It made a good diagnostic tool here.

forclip · 10th July 2012, 08:14

Great. Now I need to get a new build to check how it works..

tebasuna51 · 10th July 2012, 22:02

Quote:

Originally Posted by IanB

...
EnsureVBRMP3Sync() adds a 1MB audio cache and causes a high penalty for any out of order accesses outside the audio cache (a seek to zero plus a linear read up to the new position). It made a good diagnostic tool here.

The: (a seek to zero plus a linear read up to the new position)

is only needed for VBR audio, with CBR you can access directly to needed frame to decode.

The GetAudio function from NicAc3Source do (with SAMPLES_PER_FRAME, StreamOffset and FrameLength fix and know for each ac3):

Quote:

if (start < BufferStart || start > BufferEnd)
{
// Calculate the index and the sample range of the required frame
_FrameIndex = FrameIndex;
FrameIndex = int(start / SAMPLES_PER_FRAME);
BufferStart = __int64(FrameIndex) * SAMPLES_PER_FRAME;
BufferEnd = BufferStart + SAMPLES_PER_FRAME;

// Seek to the frame, if required
if (FrameIndex != _FrameIndex + 1)
{
_fseeki64(Stream, StreamOffset + __int64(FrameIndex) * __int64(FrameLength), SEEK_SET);

forclip · 10th July 2012, 23:52

IanB
Thanks for the fix and thanks for the build, seems like this issue is not reproducible anymore

tebasuna51 · 11th July 2012, 12:10

Quote:

Originally Posted by forclip

IanB
Thanks for the fix and thanks for the build, seems like this issue is not reproducible anymore

What build work for you?
The last AVS 2.6.0 Alpha 3 [110525] in SourceForge or a new one?

IanB · 11th July 2012, 23:19

I gave forclip a private copy of my experimental development version. I have checked the fix into CVS.

The problem was as described above with the ac_expected_next calc being to late in the code.

Wilbert · 29th July 2012, 15:46

Hi, two comments:

In conditional_functions.cpp:

Code:

AVSValue MinMaxPlane::MinMax(AVSValue clip, void* user_data, float threshold, int plane, int mode, IScriptEnvironment* env) {
  ...
    env->ThrowError("Compare Plane: This filter can only be used within run-time filters");

should be MinMax Plane instead of Compare Plane.

Also in avisynth.cpp:

Code:

PVideoFrame __stdcall ScriptEnvironment::NewVideoFrame(const VideoInfo& vi, int align) {
  ...

  if (vi.IsPlanar() && !vi.IsY8()) { // Planar requires different math ;)
    const int xmod  = 1 << vi.GetPlaneWidthSubsampling (PLANAR_U);

empty space should be removed (between sampling and the bracket), or is that allowed?

Gavino · 29th July 2012, 16:32

Quote:

Originally Posted by Wilbert

Code:

    const int xmod  = 1 << vi.GetPlaneWidthSubsampling (PLANAR_U);

empty space should be removed (between sampling and the bracket), or is that allowed?

White space is not significant, so it's OK.
But perhaps for aesthetic reasons, all three uses of GetPlaneXXXSubsampling in that function should be written the same way.

StainlessS · 29th July 2012, 20:26

This looks like a bug from Y8 to ConvertToRGB24(), to me.

Code:

Colorbars().convertToY8()
Addborders(0,0,5,0)
#crop(0,0,-5,0)		# Uncomment, No Red Stripe.
ConvertToRGB24()	# Produces LHS vertical Red Stripe in Current VDub & MediaPlayer Classic Home Cinema.
#ConvertToRGB32() 	# No red stripe
#Addborders(3,0,0,0)    # EDIT: Red stripe still there.

EDIT: Tried values up to about 15, only adding 5 on RHS produced the single pixel wide stripe.

Wilbert · 29th July 2012, 22:07

Quote:

EDIT: Tried values up to about 15, only adding 5 on RHS produced the single pixel wide stripe.

Good one. The following shows the same problem:

Code:

Colorbars(width=645, height=480, pixel_type="YV24")
ConvertToRGB24()

It looks like the left column of pixels is not being written in ConvertYV24ToRGB for pixel_step=3 (RGB24) in some cases. It's in the assembly stuff i think, so i need to pass.

StainlessS · 29th July 2012, 23:29

Code:

Colorbars().convertToY8()
ERR=1				# Vary ERR 0+, Errors at eg 5, 21, 37, 53, 69
Addborders(0,0,ERR*16 + 5,0)
#crop(0,0,-5,0)		# Uncomment, No Red Stripe.
ConvertToRGB24()	# Produces LHS Red Stripe in Current VDub & MediaPlayer Class Home Cinema.
#ConvertToRGB32() 	# No red stripe
#Addborders(3,0,0,0) # EDIT: Red stripe still there.

Seems to be every multiple of 16 + 5

IanB · 30th July 2012, 00:07

Yes the RGB24 overrun save and mask code is borked. It shows up for width%16 == 5.

:Edit: RGB24 AddBorders(N, x, 0, y), N not mod 12 or not mod 16 is also borked.

IanB · 2nd August 2012, 02:30

Fix committed to CVS ConvertToRGB24()

Fix committed to CVS AddBorders()

ryrynz · 6th August 2012, 09:33

I have to ask.. how's development coming along IanB? Your first post for this thread mentioned an update when you had more time three years ago.

StainlessS · 6th August 2012, 10:46

There is an update and its called v2.6.0 alpha3 25 May 2011 (about 14.5 months ago, not 3 years).
See 1st post.

EDIT: Thread was started 2009, but has endured several editions of v2.6.
EDIT: Perhaps a more (current) full date in thread title would be better, ie the year.

cretindesalpes · 15th August 2012, 13:10

Why don't we have IsY8 and IsYV411 at the script level? We currently have to write helper functions to achieve this, which is not very convenient. To know the chroma subsampling H & V ratios would be useful too.

StainlessS · 15th August 2012, 14:04

Quote:

Originally Posted by cretindesalpes

Why don't we have IsY8 and IsYV411 at the script level? We currently have to write helper functions to achieve this, which is not very convenient. To know the chroma subsampling H & V ratios would be useful too.

+1 on that.

9th July 2012, 01:28	#322 \| Link
forclip Registered User Join Date: Dec 2009 Posts: 63	Ok. With ColorBars() I also can reproduce this strange slowness issue, however audio output seems to be fine and it is not as slow as with the real sources. Using this script: Code: ColorBars().Trim(0, 50000) MergeChannels(last, last, last) MergeChannels(last, last) ConvertAudioTo16bit() with AviSynth 2.6 from 2009 it takes ~18sek to find the peaks. With more recent version(s) it takes ~23sek. Also, to get broken output my previous script can be simplified like this: Code: video = BlankClip(length=4248, fps=25.000, color=$000000) audio = bassAudioSource("E:\Temp\1804_0.ac3")#.EnsureVBRMP3Sync() AudioDub(video, audio) MergeChannels(last, last) ConvertAudioTo16bit() The result is attached (and I will PM you with the source file). Uncommenting EnsureVBRMP3Sync() makes the things SLOW AS HELL, it takes something aboud 15 minutes to process! However, the output seems to be OK, but I've tried it only once since it is really VERY-VERY slow. But as I mentioned in my first post, this slowness sometimes appears, sometimes not, but most of the time it appears. Replacing avisynth.dll to version from 2009 - and exactly the same script ALWAYS takes only 4-5sek to precess, w\o any slowness. Before switching to AviSynth 2.6 (from 2.5.8 MT) I've never heard from my users about any problems related to audio processing (except "audio crash" that in 2.6 is fixed), now after we are using AviSynt 2.6, some of them reporting me that audio processing is SLOW from now. I can reproduce it on my side, also I can reproduce it in MeGUI, so this is not my app fault. This strange slowness can be reproduced with avs2wav too, but output seems to be OK, or I've tested it not very well.. Any ideas? Attached Images

9th July 2012, 09:12	#323 \| Link
tebasuna51 Moderator Join Date: Feb 2005 Location: Spain Posts: 6,915	Please, send me also the source file 1804_0.ac3 to see if I can detect any trouble decoding whit NicAc3Source/bassAudioSource. I don't know for what EnsureVBRMP3Sync() can make differences decoding CBR audio, like ac3 must be. __________________ BeHappy, AviSynth audio transcoder.

10th July 2012, 05:02	#324 \| Link
IanB Avisynth Developer Join Date: Jan 2003 Location: Melbourne, Australia Posts: 3,167	@forclip, Between 2.6.0.1 and 2.6.0.2 the ac_expected_next update line was moved later in the code, so for the no audio cache state the skipping chunks case was constantly being incorrectly triggered. This led to all the default audio caches needlessly becoming enabled resulting in lots of extra memcpy'ing happening, slowing things down. Not sure if it's the whole cause but it does have quite a performance impact. Code: ... long _cs = ac_currentscore; if (start < ac_expected_next) _cs = InterlockedExchangeAdd(&ac_currentscore, -5) - 5; // revisiting old ground - a cache could help else if (start > ac_expected_next) _cs = InterlockedDecrement(&ac_currentscore); // skiping chunks - a cache might not help else // (start == ac_expected_next) _cs = InterlockedIncrement(&ac_currentscore); // strict linear reading - why bother with a cache ... >> ac_expected_next = start + count; if (h_audiopolicy == CACHE_AUDIO_NONE \|\| h_audiopolicy == CACHE_AUDIO_NOTHING) { child->GetAudio(buf, start, count, env); return; // We are ok to return now! } ... << ac_expected_next = start + count; while (count>maxsamplecount) { //is cache big enough? @tebasuna51, EnsureVBRMP3Sync() adds a 1MB audio cache and causes a high penalty for any out of order accesses outside the audio cache (a seek to zero plus a linear read up to the new position). It made a good diagnostic tool here.

29th July 2012, 20:26	#332 \| Link
StainlessS HeartlessS Usurer Join Date: Dec 2009 Location: Over the rainbow Posts: 10,980	This looks like a bug from Y8 to ConvertToRGB24(), to me. Code: Colorbars().convertToY8() Addborders(0,0,5,0) #crop(0,0,-5,0) # Uncomment, No Red Stripe. ConvertToRGB24() # Produces LHS vertical Red Stripe in Current VDub & MediaPlayer Classic Home Cinema. #ConvertToRGB32() # No red stripe #Addborders(3,0,0,0) # EDIT: Red stripe still there. EDIT: Tried values up to about 15, only adding 5 on RHS produced the single pixel wide stripe. __________________ I sometimes post sober. StainlessS@MediaFire ::: AND/OR ::: StainlessS@SendSpace "Some infinities are bigger than other infinities", but how many of them are infinitely bigger ??? Last edited by StainlessS; 29th July 2012 at 21:05.

29th July 2012, 23:29	#334 \| Link
StainlessS HeartlessS Usurer Join Date: Dec 2009 Location: Over the rainbow Posts: 10,980	Code: Colorbars().convertToY8() ERR=1 # Vary ERR 0+, Errors at eg 5, 21, 37, 53, 69 Addborders(0,0,ERR16 + 5,0) #crop(0,0,-5,0) # Uncomment, No Red Stripe. ConvertToRGB24() # Produces LHS Red Stripe in Current VDub & MediaPlayer Class Home Cinema. #ConvertToRGB32() # No red stripe #Addborders(3,0,0,0) # EDIT: Red stripe still there. Seems to be every multiple of 16 + 5 __________________ I sometimes post sober. StainlessS@MediaFire ::: AND/OR :::* StainlessS@SendSpace "Some infinities are bigger than other infinities", but how many of them are infinitely bigger ???

8th July 2012, 11:36	#321 \| Link
IanB Avisynth Developer Join Date: Jan 2003 Location: Melbourne, Australia Posts: 3,167	@forclip, With the script as quoted, and substituting a 384Kbit stereo .AC3 file I had lying around, I see no problem with any versions of Avisynth.dll from 2.5.8.5, 2.6.0.0, 2.6.0.1, 2.6.0.2, current CVS and my current development version. I used a fairly old modified NicAudio.dll I built myself based on version 2.0.0.5. circa 2007. I tried a single pass linear save of the output stream as well as random spot playback. From your avs_audio.png image, the bass25.wav sample has a big jump in level at around 2:10, this might be bugs with decoding your file with seeking. Add the EnsureVBRMP3Sync() filter to your NicAC3Source() and observe any change in behaviour. Also try to reproduce the results using ColorBars() in place of NicAC3Source() for the audio source.

10th July 2012, 08:14	#325 \| Link
forclip Registered User Join Date: Dec 2009 Posts: 63	Great. Now I need to get a new build to check how it works..

10th July 2012, 23:52	#327 \| Link
forclip Registered User Join Date: Dec 2009 Posts: 63	IanB Thanks for the fix and thanks for the build, seems like this issue is not reproducible anymore

11th July 2012, 23:19	#329 \| Link
IanB Avisynth Developer Join Date: Jan 2003 Location: Melbourne, Australia Posts: 3,167	I gave forclip a private copy of my experimental development version. I have checked the fix into CVS. The problem was as described above with the ac_expected_next calc being to late in the code.

30th July 2012, 00:07	#335 \| Link
IanB Avisynth Developer Join Date: Jan 2003 Location: Melbourne, Australia Posts: 3,167	Yes the RGB24 overrun save and mask code is borked. It shows up for width%16 == 5. :Edit: RGB24 AddBorders(N, x, 0, y), N not mod 12 or not mod 16 is also borked. Last edited by IanB; 31st July 2012 at 06:43.

2nd August 2012, 02:30	#336 \| Link
IanB Avisynth Developer Join Date: Jan 2003 Location: Melbourne, Australia Posts: 3,167	Fix committed to CVS ConvertToRGB24() Fix committed to CVS AddBorders()

6th August 2012, 09:33	#337 \| Link
ryrynz Registered User Join Date: Mar 2009 Posts: 3,650	I have to ask.. how's development coming along IanB? Your first post for this thread mentioned an update when you had more time three years ago.

6th August 2012, 10:46	#338 \| Link
StainlessS HeartlessS Usurer Join Date: Dec 2009 Location: Over the rainbow Posts: 10,980	There is an update and its called v2.6.0 alpha3 25 May 2011 (about 14.5 months ago, not 3 years). See 1st post. EDIT: Thread was started 2009, but has endured several editions of v2.6. EDIT: Perhaps a more (current) full date in thread title would be better, ie the year. __________________ I sometimes post sober. StainlessS@MediaFire ::: AND/OR ::: StainlessS@SendSpace "Some infinities are bigger than other infinities", but how many of them are infinitely bigger ??? Last edited by StainlessS; 6th August 2012 at 17:32.

15th August 2012, 13:10	#339 \| Link
cretindesalpes ͡҉҉ ̵̡̢̛̗̘̙̜̝̞̟̠͇̊̋̌̍̎̏̿̿ Join Date: Feb 2009 Location: No support in PM Posts: 712	Why don't we have IsY8 and IsYV411 at the script level? We currently have to write helper functions to achieve this, which is not very convenient. To know the chroma subsampling H & V ratios would be useful too. __________________ dither 1.28.1 for AviSynth \| avstp 1.0.4 for AviSynth development \| fmtconv r30 for Vapoursynth & Avs+ \| trimx264opt segmented encoding

Thread Tools	Search this Thread
Show Printable Version Email this Page	Search this Thread: Advanced Search
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode