LAME: different compilers producing significantly different bitstreams, Split from “LAME 3.99 is out” Options

[mixminus1]

No, latest 64-bit version of 3.99.4 (20120125) from Rarewares has higher bitrates on the quieter track, as well, although not quite as much as the 32-bit version

Encodes done on same machine, but under Win7 64-bit, again from the command line, and I've included the bitrates from the previous test:

CODE

Shaolin Spirit (duo)

3.99.4 Rarewares 32-bit @ V2: 170.3 kbps
3.99.4 Rarewares 64-bit @ V2: 168.7 kbps
3.99.4 lvqcl 32-bit @ V2: 159.8 kbps

3.99.4 Rarewares 32-bit @ V5: 104.8 kbps
3.99.4 Rarewares 64-bit @ V5: 103.7 kbps
3.99.4 lvqcl 32-bit @ V5: 95.1

Rolodex Propaganda

3.99.4 Rarewares 32-bit @ V2: 200.1 kbps
3.99.4 Rarewares 64-bit @ V2: 200.0 kbps
3.99.4 lvqcl 32-bit @ V2: 199.3 kbps

3.99.4 Rarewares 32-bit @ V5: 135.8 kbps
3.99.4 Rarewares 64-bit @ V5: 135.8 kbps
3.99.4 lvqcl 32-bit @ V5: 135.6 kbps

[john33]

No, latest 64-bit version of 3.99.4 (20120125) from Rarewares has higher bitrates on the quieter track, as well, although not quite as much as the 32-bit version

Encodes done on same machine, but under Win7 64-bit, again from the command line, and I've included the bitrates from the previous test:

...

Those differences are entirely within what I would expect as the 64 bit compiles do not use the nasm code, although the VC10 numbers still differ more than I would expect. I'll compile using VC10 later and see what differences I get.

This post has been edited by john33: Jan 31 2012, 09:37

[john33]

OK, done some testing comparing the Rarewares 32 bit compile (ICL12.1) against my own VC10 32 bit compile (not ICL). Results are quite interesting (I think!!):

CODE

Microsoft Windows [Version 6.1.7601]
Copyright © 2009 Microsoft Corporation. All rights reserved.

F:\testdir>lame-vc9 -V 3 14.wav 14.mp3
LAME 3.99.4 32bits (http://lame.sf.net)
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used), SSE2
Using polyphase lowpass filter, transition band: 17960 Hz - 18494 Hz
Encoding 14.wav to 14.mp3
Encoding as 44.1 kHz j-stereo MPEG-1 Layer III VBR(q=3)
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
14177/14177 (100%)| 0:10/ 0:10| 0:10/ 0:10| 36.978x| 0:00
32 [ 1] *
40 [ 0]
48 [ 18] %
56 [ 630] %%%*****
64 [ 2172] %%%%%%%%%%%%**************
80 [ 78] %
96 [ 28] %
112 [ 39] %
128 [ 1081] %************
160 [ 5740] %%%%%%%%%%%%%%%%%%%%%%%%%%%%***************************************
192 [ 2469] %%%%%%%%%%%%%%%%%%***********
224 [ 454] %%%***
256 [ 1036] %%%%%%%******
320 [ 431] %%%***
-------------------------------------------------------------------------------
kbps LR MS % long switch short %
156.9 44.7 55.3 85.4 7.9 6.7
Writing LAME Tag...done
ReplayGain: 0.0dB

F:\testdir>lame-vc10 -V 3 14.wav 14.mp3
LAME 3.99.4 32bits (http://lame.sf.net)
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used), SSE2
Using polyphase lowpass filter, transition band: 17960 Hz - 18494 Hz
Encoding 14.wav to 14.mp3
Encoding as 44.1 kHz j-stereo MPEG-1 Layer III VBR(q=3)
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
14177/14177 (100%)| 0:10/ 0:10| 0:10/ 0:10| 35.327x| 0:00
32 [ 4] %
40 [ 35] %
48 [ 1022] %%%%%*******
56 [ 1640] %%%%%%%%%***********
64 [ 184] %%*
80 [ 25] %
96 [ 36] %
112 [ 30] %
128 [ 1050] %************
160 [ 5768] %%%%%%%%%%%%%%%%%%%%%%%%%%%%***************************************
192 [ 2477] %%%%%%%%%%%%%%%%%%%**********
224 [ 443] %%%***
256 [ 1032] %%%%%%%*****
320 [ 431] %%%***
-------------------------------------------------------------------------------
kbps LR MS % long switch short %
155.0 44.7 55.3 85.4 7.9 6.7
Writing LAME Tag...done
ReplayGain: 0.0dB

F:\testdir>lame-vc9 -V 5 14.wav 14.mp3
LAME 3.99.4 32bits (http://lame.sf.net)
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used), SSE2
Using polyphase lowpass filter, transition band: 16538 Hz - 17071 Hz
Encoding 14.wav to 14.mp3
Encoding as 44.1 kHz j-stereo MPEG-1 Layer III VBR(q=5)
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
14177/14177 (100%)| 0:09/ 0:09| 0:09/ 0:09| 38.541x| 0:00
32 [ 23] %
40 [ 1025] %%%*********
48 [ 1755] %%%%%%**************
56 [ 86] %
64 [ 28] %
80 [ 42] %
96 [ 302] %***
112 [ 1986] %%********************
128 [ 6081] %%%%%%%%%%%%%%%%%%%%%%%%%%%%%**************************************
160 [ 1267] %%%%%%%%%*****
192 [ 684] %%%*****
224 [ 500] %%%***
256 [ 366] %%***
320 [ 32] %
-------------------------------------------------------------------------------
kbps LR MS % long switch short %
121.0 34.6 65.4 85.4 7.9 6.7
Writing LAME Tag...done
ReplayGain: 0.0dB

F:\testdir>lame-vc10 -V 5 14.wav 14.mp3
LAME 3.99.4 32bits (http://lame.sf.net)
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used), SSE2
Using polyphase lowpass filter, transition band: 16538 Hz - 17071 Hz
Encoding 14.wav to 14.mp3
Encoding as 44.1 kHz j-stereo MPEG-1 Layer III VBR(q=5)
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
14177/14177 (100%)| 0:10/ 0:10| 0:10/ 0:10| 36.358x| 0:00
32 [ 2843] %%%%%%%%************************
40 [ 40] %
48 [ 8] %
56 [ 16] %
64 [ 32] %
80 [ 37] %
96 [ 316] %***
112 [ 1977] %%********************
128 [ 6070] %%%%%%%%%%%%%%%%%%%%%%%%%%%%%**************************************
160 [ 1265] %%%%%%%%%*****
192 [ 669] %%%*****
224 [ 510] %%%***
256 [ 357] %%**
320 [ 37] %
-------------------------------------------------------------------------------
kbps LR MS % long switch short %
118.2 34.6 65.4 85.4 7.9 6.7
Writing LAME Tag...done
ReplayGain: 0.0dB

F:\testdir>lame-vc9 -V 7 14.wav 14.mp3
LAME 3.99.4 32bits (http://lame.sf.net)
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used), SSE2
Resampling: input 44.1 kHz output 32 kHz
polyphase lowpass filter disabled
Encoding 14.wav to 14.mp3
Encoding as 32 kHz j-stereo MPEG-1 Layer III VBR(q=5.63333)
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
10287/10287 (100%)| 0:09/ 0:09| 0:09/ 0:09| 37.801x| 0:00
32 [ 1184] %%%%%%%******************
40 [ 905] %%%%%%*************
48 [ 20] %
56 [ 16] %
64 [ 71] %*
80 [ 1216] %%************************
96 [ 3250] %%%%%%%%%%%%%%%%%%%%%%%%%%*****************************************
112 [ 1811] %%%%%%%%%%%%%%%%%%%%%%****************
128 [ 783] %%%%%%%%*********
160 [ 627] %%%%%********
192 [ 230] %%***
224 [ 152] %%**
256 [ 21] %
320 [ 1] *
-------------------------------------------------------------------------------
kbps LR MS % long switch short %
95.0 36.1 63.9 86.1 7.8 6.2
Writing LAME Tag...done
ReplayGain: +0.1dB

F:\testdir>lame-vc10 -V 7 14.wav 14.mp3
LAME 3.99.4 32bits (http://lame.sf.net)
CPU features: MMX (ASM used), 3DNow! (ASM used), SSE (ASM used), SSE2
Resampling: input 44.1 kHz output 32 kHz
polyphase lowpass filter disabled
Encoding 14.wav to 14.mp3
Encoding as 32 kHz j-stereo MPEG-1 Layer III VBR(q=5.63333)
Frame | CPU time/estim | REAL time/estim | play/CPU | ETA
10287/10287 (100%)| 0:10/ 0:10| 0:10/ 0:10| 35.066x| 0:00
32 [ 2096] %%%%%%%%%%%%%*******************************
40 [ 12] %
48 [ 19] %
56 [ 11] %
64 [ 70] %*
80 [ 1217] %%************************
96 [ 3254] %%%%%%%%%%%%%%%%%%%%%%%%%%*****************************************
112 [ 1820] %%%%%%%%%%%%%%%%%%%%%%****************
128 [ 755] %%%%%%%%********
160 [ 634] %%%%%*********
192 [ 220] %%***
224 [ 157] %%**
256 [ 21] %
320 [ 1] *
-------------------------------------------------------------------------------
kbps LR MS % long switch short %
94.1 36.1 63.9 86.1 7.8 6.2
Writing LAME Tag...done
ReplayGain: +0.1dB

F:\testdir>

This track is neither particularly quiet, nor loud, but what is interesting is that the very quiet parts seem to attract a higher bit allocation with the Intel compile than with the VC10 compile. So, I'm guessing that with particularly quiet tracks, this accounts for the larger difference between the different compiles.

[Pretentiousman]

Is it possible for an executable from one compiler to produce a difference in quality or sound than another executable from different compiler? It seems that a difference in bitrate is possible. I am curious. I am no programmer btw.

This post has been edited by Pretentiousman: Jan 31 2012, 18:00

[lvqcl]

2 seconds of Liszt.wav (LAME -V1):

original track, then my compile, then msvs2010 debug compile, then compile from Rarewares.




And, differences between original and encoded files: my / my msvs2010 debug / Rarewares compile.



bitrates, kbps: 207(my,msvs) / 214(debug) / 246(rarewares).

This post has been edited by lvqcl: Jan 31 2012, 18:10

[john33]

Is it possible for an executable from one compiler to produce a difference in quality or sound than another executable from different compiler? It seems that a difference in bitrate is possible. I am curious. I am no programmer btw.

A question that has been asked many times. The short answer is that I don't believe anyone has claimed to hear any verifiable discernable differences and that is all that matters.

[JJZolx]

Is it possible for an executable from one compiler to produce a difference in quality or sound than another executable from different compiler? It seems that a difference in bitrate is possible.

Which, IMO, is no less concerning than if they produced differences in sound quality.

The short answer is that I don't believe anyone has claimed to hear any verifiable discernable differences and that is all that matters.

It's not all that matters.

The whole point of a lossy codec using VBR is to produce the lowest possible bitrate at a given SQ level. Perhaps it's naïve, but I would expect all compiles on all platforms to produce the same encoding results.

This post has been edited by JJZolx: Jan 31 2012, 18:19

[lvqcl]

Perhaps it's naïve, but I would expect all compiles on all platforms to produce the same encoding results.

LAME uses floating point calculations, so it is highly unlikely.

[JJZolx]

But to what degree in terms of overall bitrate? These results seem to be all over the map.

[Alex B]

Here's how EncSpot Pro sees the differences in the -V5 encoded "Liszt" samples.

lvqcl's compile:


The fixed Rarewares compile:


Bit graph (bitrate/time), both compiles in an animated gif:


I have no idea if the differences can be audible.

This post has been edited by Alex B: Jan 31 2012, 18:58

[apodtele]

Perhaps it's naïve, but I would expect all compiles on all platforms to produce the same encoding results.

LAME uses floating point calculations, so it is highly unlikely.

Please... Given the same CPU, floating point calculations are quite deterministic.

The whole story smells like LAME is too careless about floating overflows and underflows.

[C.R.Helmrich]

Please... Given the same CPU, floating point calculations are quite deterministic.

I wish... well, yes, maybe the calculations are deterministic, but the order of calculations (which can vary between compilers according to optimization algorithms) can have a significant influence on the result. But of course the observations discussed here are extreme.

Chris

[apodtele]

Order is defined by precedence and compiler rarely dares to go against left-to-right as specified by the C standard.
I still bet on undeflows of the type that 3.99.4 was trying to paper over with max(x, epsilon). This type of bugfixing is dangerous.

This post has been edited by db1989: Feb 7 2012, 14:43

Reason for edit: removing unnecessary full quote of above post

[JJZolx]

Here's some more data at different VBR quality levels with the fixed RareWares 32-bit compile of 3.99.4. Perhaps it's a little more real world than the 30 second Liszt piano solo (which is included), as all of these tracks are at least several minutes long. These bitrates are more consistent, although lvqcl's compile consistently comes out ahead. The Thelonious Monk track is also a piano solo piece. Interestingly, it shows probably the biggest difference between the two compiles after the Liszt sample.

(My apologies if this is too much data for the forums. It doesn't appear that I can attach a simple text log.) db1989: You can paste the log into a codebox (rather than between code tags) or upload it to our dedicated Uploads subforum. Editing according to the latter:

CODE

Aerosmith - Sweet Emotion.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 269 228 190 168 149 129 113 100 83
3.99.4 (lvqcl) 264 223 188 168 149 129 112 99 83

Allman Brothers - Melissa.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 263 231 202 184 164 142 124 112 90
3.99.4 (lvqcl) 259 227 198 184 164 142 123 111 90

Bill Evans - Polka Dots And Moonbeams.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 243 207 178 164 144 124 107 97 80
3.99.4 (lvqcl) 243 205 176 164 144 124 105 96 80

Black Keys - Set You Free.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 259 222 186 165 146 128 115 98 79
3.99.4 (lvqcl) 258 219 185 165 146 128 115 98 79

Bob Marley - Redemption Song.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 241 211 175 151 129 107 93 82 72
3.99.4 (lvqcl) 227 197 167 151 128 103 86 78 72

Count Basie - Lester Leaps In.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 280 244 199 180 158 135 120 107 93
3.99.4 (lvqcl) 257 228 196 179 156 130 112 102 93

Cowboy Junkies - Mining For Gold.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 249 208 176 159 139 120 104 93 77
3.99.4 (lvqcl) 248 206 174 159 139 119 102 93 77

Del McCoury - Rambling Boy.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 275 238 204 177 155 131 113 104 90
3.99.4 (lvqcl) 266 225 192 177 155 130 110 102 90

Dizzy Gillespie - Manteca.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 266 229 198 181 162 139 121 109 94
3.99.4 (lvqcl) 262 223 195 181 162 139 119 108 94

Ella Fitzgerald - Anything Goes.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 187 162 142 125 112 97 86 75 63
3.99.4 (lvqcl) 187 159 137 123 109 92 79 71 60

George Baker Selection - Little Green Bag.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 268 236 206 190 171 150 133 120 98
3.99.4 (lvqcl) 267 235 205 190 171 150 133 120 98

George Jones - She Thinks I Still Care.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 258 209 179 164 146 127 111 102 87
3.99.4 (lvqcl) 258 206 177 164 146 126 109 102 87

Jack Johnson - Better Together.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 264 229 195 172 152 130 114 103 87
3.99.4 (lvqcl) 262 226 192 172 152 130 113 103 87

Jimmy Cliff - Ska All Over The World.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 258 228 195 179 160 134 115 106 93
3.99.4 (lvqcl) 248 217 190 179 160 133 113 105 93

John Coltrane - Giant Steps.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 291 241 205 187 167 145 127 121 102
3.99.4 (lvqcl) 291 236 199 187 167 145 126 120 102

Johnny Winter - Mojo Boogie.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 275 238 200 169 149 128 112 97 79
3.99.4 (lvqcl) 259 221 187 169 149 128 110 96 79

Kanye West - Gold Digger.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 249 214 182 160 144 127 112 97 77
3.99.4 (lvqcl) 247 211 176 159 143 125 108 96 77

Katia Lebeque & the Cleveland Orchestra - Rhapsody in Blue.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 267 234 196 161 140 116 102 91 86
3.99.4 (lvqcl) 247 210 178 159 135 108 91 83 82

Leo Kottke - Jesus Maria.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 240 210 185 165 145 121 103 96 87
3.99.4 (lvqcl) 236 204 177 165 144 118 98 94 86

Liszt.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 283 245 183 156 136 119 111 103 101
3.99.4 (lvqcl) 257 207 172 149 125 97 82 76 83

Lotus - Sea Of Tranquility.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 231 196 166 145 128 110 96 86 73
3.99.4 (lvqcl) 228 191 160 142 124 103 88 79 69

Marcos Valle - Para Os Filhos De Abraao.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 261 222 187 167 148 128 114 101 82
3.99.4 (lvqcl) 260 220 186 167 148 128 113 101 81

Nirvana - Smells Like Teen Spirit.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 260 227 193 169 148 128 114 97 79
3.99.4 (lvqcl) 258 223 190 169 148 127 113 96 79

Ottmar Liebert - August Moon.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 300 255 222 202 182 158 136 130 107
3.99.4 (lvqcl) 298 246 216 202 181 157 134 129 107

Peter Tosh - Get Up, Stand Up.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 281 251 220 200 178 152 134 124 107
3.99.4 (lvqcl) 276 246 216 200 177 151 131 123 106

Pointer Sisters - Fire.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 264 230 195 169 149 128 112 102 89
3.99.4 (lvqcl) 261 225 191 169 148 126 110 101 89

Psychedelic Furs - Love My Way.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 267 225 183 164 145 126 111 98 81
3.99.4 (lvqcl) 261 219 181 164 145 125 110 97 81

Rolling Stones - Street Fighting Man.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 270 235 206 192 172 148 123 112 90
3.99.4 (lvqcl) 259 231 204 192 171 147 119 112 90

Shirley Bassey - Moonraker.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 264 229 198 178 157 136 119 103 81
3.99.4 (lvqcl) 261 225 194 177 157 135 118 103 80

Smiths - Please, Please, Please, Let Me Get What I Want.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 271 246 217 195 173 150 133 105 78
3.99.4 (lvqcl) 271 246 217 195 173 150 132 105 78

Steely Dan - Black Friday.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 266 231 196 174 152 129 115 103 85
3.99.4 (lvqcl) 266 231 195 174 152 128 113 102 85

Stevie Ray Vaughan - The House Is Rockin'.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 262 231 201 170 150 130 115 101 83
3.99.4 (lvqcl) 261 229 199 170 150 130 115 100 83

Sublime - Badfish.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 280 242 204 182 161 139 123 113 94
3.99.4 (lvqcl) 275 237 201 182 161 139 122 113 94

Thelonious Monk - Ruby, My Dear.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 229 201 174 153 135 110 94 86 85
3.99.4 (lvqcl) 226 196 166 150 132 101 80 75 80

Tomita - Introduction And Dance Of The Firebird.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 250 225 197 165 141 113 97 86 81
3.99.4 (lvqcl) 236 209 182 164 139 107 88 79 78

U2 - Bullet The Blue Sky.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 256 224 192 168 148 127 113 99 79
3.99.4 (lvqcl) 255 222 190 168 148 127 113 99 79

Van Morrison - Moondance.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 259 225 193 174 155 133 116 105 87
3.99.4 (lvqcl) 254 218 188 174 155 133 114 105 87

Will Ackerman - Abide The Winter.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 255 230 200 168 143 113 95 87 80
3.99.4 (lvqcl) 244 215 186 167 141 108 88 82 78

Yngwie Malmsteen - Leviathan.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 271 228 189 167 146 127 112 96 81
3.99.4 (lvqcl) 268 222 185 166 146 126 111 96 81

Yo-Yo Ma - Boccherini - Cello Concerto No. 7 in G major, G480 I. Allegro.wav
-V0 -V1 -V2 -V3 -V4 -V5 -V6 -V7 -V8
3.99.4 (RareWares) 250 223 189 165 143 118 101 91 83
3.99.4 (lvqcl) 239 209 180 164 141 114 95 87 82

This post has been edited by db1989: Feb 7 2012, 14:39

Reason for edit: see note in green

[saratoga]

Perhaps it's naïve, but I would expect all compiles on all platforms to produce the same encoding results.

LAME uses floating point calculations, so it is highly unlikely.

Please... Given the same CPU, floating point calculations are quite deterministic.

Given the same CPU and the same binary, they usually are. But across different platforms or different compilers, they are not.

Order is defined by precedence and compiler rarely dares to go against left-to-right as specified by the C standard.

Its not that simple:

http://msdn.microsoft.com/en-us/library/aa...7(v=vs.71).aspx

c specifies operator precedence, but that does not strictly define order, only how a statement is parsed. Compilers still (often quite aggressively) reorder floating point calculations depending on the optimization level and compiler flags. You can of course disable this, but the performance penalty can be quite large, particularly with vector extensions.

When talking about different instructions sets, pretty much all bets are off. Operations can be done at different precision and in different order.

[nu774]

CODE

main()
{
    volatile double x = 53, y = x / 100.0 * 100.0;
    volatile int ny = x / 100.0 * 100.0;
    int nyy = y;
    printf("%g %d %d\n", y, ny, nyy);
}

In my environment, i686-w64-mingw32-gcc and cygwin gcc prints "53 52 53" (On the other hand, CL compiler of MSVC10 prints "53 53 53").
The difference between ny and nyy is only whether calculated value (x / 100.0 * 100.0) is once stored to a double variable y or not.
x86 architecture has fp register wider than double precision(64bit). Therefore, just a store/load can make such difference (introduced by rounding error).

[apodtele]

54 52 53

(introduced by rounding error).

You should add 0.5 to achieve proper rounding with whatever compiler.
We need a proper investigation rather than brushing it under ABX rug.

[nu774]

You should add 0.5 to achieve proper rounding with whatever compiler.

Or, just use lrint()
However, even if you use "proper rounding", it still remains the same that a simple load/store can change the value of floating point, at least on some architecture.
My point was, how operations (such as load/store) are used/ordered is beyond control of a programmer who programs in high level language such as C, and also varies with compilers or optimization settings or something.

We need a proper investigation rather than brushing it under ABX rug.

Agree with it.

This post has been edited by nu774: Feb 1 2012, 04:53

[nu774]

We need a proper investigation rather than brushing it under ABX rug.

Agree with it.

O My English...
I don't know if it is in the range of expected result of floating point math error. If it is, I don't think investigation should necessarily be done.
Probably Lame dev can answer it.

[timcupery]

We need a proper investigation rather than brushing it under ABX rug.

Very well-said. (not that we can require anyone to do an investigation - all of us LAME users are in debt to the devs as with anything open-source)
But marked differences in bitrate are weirdly striking.


What I wonder: if you ran the differing-bitrate files through mp3packer, would they come out with the same bitrate? i.e., are there wasted bits that can be compressed out? Worth checking at least, unless someone with knowledge of the innards of LAME can say that this wouldn't be the case.
I can run the respective files through mp3packer if anyone wants to send me the files.

[lvqcl]

Test signal: sine (-12 dBFS) + white noise (-N dBFS), where N = 3...150.

Here is a graph: MP3 bitrate as a function of N, for different compiles.

[halb27]

Looks like your MSVC10 compile behaves more reasonable - I guess the sine signal is still encoded fine.

This post has been edited by halb27: Feb 2 2012, 17:48

[lvqcl]

Spectrograms of sine+noise (3 seconds from an mp3 file encoded with MSVS compile, then Debug compile, then ICL/Rarewares compile):

N=80:



N=90:



N=95:



N=100:



N=110:

[apodtele]

Spectrograms of sine+noise (3 seconds from an mp3 file encoded with MSVS compile, then Debug compile, then

Why does this junk keeps piling up in later seconds?
Could it be uninitialized variable somewhere that is properly initially set to zero by one compiler but not by the other?

[halb27]

Looks like it's the decision process for deciding upon 'digital silence' that is varying with compiler.

lvqcl, is the white noise inaudible for N=95?

This post has been edited by halb27: Feb 2 2012, 20:36