Alpha Weighting in RGBA Image Squared Distortion Measure

I'm going to look at how the Alpha channel should be weighted vs the RGB components when measuring distortions in RGBA textures, when that RGBA texture is used for alpha blending.

(aside: of course many RGBA 4-channel images are not actually using A for opacity in alpha blending and this discussion does not apply to those; if your RGBA image is really 4 scalar channels of the same type of signal then the weight of each channel should be the same)

For concreteness, an RGBA image used for alpha blending means that we construct an output pixel like :

out = A * RGB + (1 - A) * background

RGBA in [0,1] scale here

and for most of this discussion I will assume that the distortion measure we are looking at is weighted squared difference (L2 norm) :

' = prime = indicates values in the modified texture

SD(RGB) = ( RGB' - RGB )^2 = (R' - R)^2 + (G' - G)^2 + (B' - B)^2

SD(RGBA) = SD(RGB) + (A' - A)^2

WSD(RGBA) = SD(RGB) + W(A) * (A' - A)^2

W(A) is the weight of the A term in the RGBA squared difference

In the end what we care about is what the user sees, which are the "out" pixel values after alpha blending. So the correct error metric (or weight for A) should be determined by minimizing the error in the output pixel.

(choosing a weight for A is equivalent to choosing a bit allocation ratio between RGB color channels and the A channel; in an RD encoder, changing W(A) corresponds to transferring bits)

The output pixel is RGB only (no A), and we'll assume that we want to minimize the squared difference of the output pixel. So the question is how should that be expressed in an error metric on the RGBA texture.

There are three intuitive factors which should obviously affect this issue, and we will see them come out in the math. They are :

1. When background RGB is nearly equal to the texture RGB, A doesn't matter at all. If background == RGB , then any A value produces the same output pixel. The bigger difference there is between background and the current texture, the more important A is.

2. When A is smaller, the texture RGB is less important. In the extreme case of A=0, the texture RGB doesn't matter at all, it isn't used and is fully redundant. This is a well known problem in texture compression and is why we prefer premultiplied alpha (see more later).

3. Ignoring those factors, A affects the output color in all 3 channels, so will be something like ~3 times more important than each RGB channel.

So, without further ado let's do the simple math :

E = (out' - out)^2

out = A * RGB + (1 - A) * background

let background = RGB + D

(background is an RGB vector; D is a scalar difference on each channel)

out = A * RGB + (1 - A) * (RGB + D) = RGB + (1 - A)*D

out(R) = R + (1-A)*D

out' = A' * RGB' + (1 - A') * (RGB + D)

(note the background RGB in out' is not RGB' ; the background color stays the same)

R' = R + dR
A' = A + dA

out'(R) = (A + dA) * (R + dR) + (1 - A-dA) * (R+D)

out'(R) = (A + dA) * dR + R + (1 - A-dA) * D

out'(R) = A * dR + R + (1 - A)*D -dA * D

(dropping terms squared in the delta d)

(out' - out)(R) = A * dR - dA * D

E = A^2 dRGB^2 + 3*D^2 dA^2 + linear terms

linear terms = - 2 * A * D * dA * dRGB

dropping the linear terms (see later on why) :

E = A^2 * dRGB^2 + 3*D^2 * dA^2

This is equivalent to a weight on alpha :

W(A) = 3 * D^2 / A^2

D = difference between texture and background (scalar)

note instead of 3*D^2 you could use the vector RGB difference to background :

W(A) = D_RGB^2 / A^2

This weight term encapsulates the three intuitive principles we saw at the beginning: when the foreground and background are the same RGB color, then A doesn't matter, W(A) ~ D^2 goes to zero. As A goes to zero, RGB doesn't matter; W(A) goes to infinity like 1/A^2 (if you prefer, the weight of RGB goes to zero like W(RGB) ~ A^2). Other than that, if A is ~ D, then W(A) is ~ 3X the RGB weight because it affects all three channels.

(revisiting the note on dropping the linear terms : several reasons why I think this is right. First of all, because these are signed linear terms, they will average out to zero when summing over pixels, assuming your distortions dRGB are random around the target value and not net biased to one side. Second of all, because we don't really want to be measuring this effect. What the linear term measures is ways that you can compensate for miss-blending the background with a wrong A by using a wrong RGB to fix it. Say you're blending white onto a black background and your A is too high, you can compensate and lower E by making RGB lower to get the right output color. When that does happen to work in our favor we just don't even want to know about that. It also assumes exact knowledge of the background.)

Now we can't assume that we know the background. We could look at the worst case, D = 1.0, blending black on white or vice versa. That's when A matters the most, and W(A) = 3 / A^2 ; in that case :

maximal difference to background, D = 1.0

W(A = 1) = 3
W(A = 0.5) = 12
W(A = 0.25) = 48

Alpha should be weighted much higher than RGB.

(note that because of interpolation we probably don't want the weight of RGB to ever go completely to zero)

But D = 1 is rare in practice. In fact in games we very often do alpha blending when D is closer to zero. For example say you're doing some alpha blended particle effects, often you blend many fire or smoke puffs on top of each other, so they are often blending onto other fire and smoke that is very similar. On many games, the RGB palette is squarely in the gray/brown domain so we expect D to be much less than 1 most of the time.

If we assume that foreground and background are both random numbers on the unit interval [0,1] , then we would have :

D = < |x - y| >  (x and y uniform random on the unit interval)

(this is simple integral and fun exercise for the reader)

D = 1/3

W(A) = (1/3) / (A^2)

W(A = 0.5) = 4/3

Now, assuming the colors are random is also clearly wrong, don't take this as "the answer", it's just another data point, perhaps more realistic than D=1, but still not using any assumption about texture and background often matching, which would make D even smaller than 1/3. If you assume the worst case (D=1) you are over-weighting A when it's not always necessary, but assuming D very small would make you under-weight A when the background is in fact very different from the texture.

Let's set that aside and revisit something we mentioned early on : premultiplied alpha.

We like premultiplied alpha in texture compression because without it you are sending totally unnecessary bits in RGB values when A=0. With some types of textures that are largely transparent this can be a huge amount of bits wasted. (some texture codecs will not encode RGB when A=0 exactly, but premultiplied also correctly down-weights the importance of RGB when A is finite but tiny; premultiplied also filters/interpolates better)

With premultiplied alpha the blending equation is :

out = RGB + (1 - A) * background

(no *A on the RGB from the texture, A is only used to multiply the background)

Let's compute what the squared difference weight W(A) should be for RGBA textures used in premultiplied alpha. For laughs I'll do it a different way.

for small deltas :

out_R = R + (1 - A) * background_R

d(out_R)  = d/dR(out_R) * dR + d/dA(out_R) * dA

d/dR(out_R) = 1
d/dA(out_R) = - background_R

d(out_R) = (dR - background_R * dA)

E = d(out)^2 = d(out_R)^2 + d(out_G)^2 + d(out_B)^2

dropping linear terms (see arguments above)

E = dR^2 + dG^2 + dB^2 + (background_R^2 + background_G^2 + background_B^2) * dA^2

E = SD(RGB) + WPM(A) * dA^2

WPM(A) = background_R^2 + background_G^2 + background_B^2

WPM(A) = background_RGB^2

WPM(A) = 3 * background^2  (assuming the same scalar "background" value on all channels)

The weight for pre-multiplied alpha is similar to the non-PM case, but without the annoying 1/A^2 in the denominator, which is a big advantage. Our weighting no longer depends on our A at all, which means the channels can be treated independently. The A weight has been taken in the "premultipy" scaling of RGB. This is just a way of saying that our intuition for premultiplied alpha was correct : using premultiplied alpha has correctly scaled the RGB component for its importance with alpha, so that you can use the L2 norm on the RGBA channels without any inter-channel correction factor.

Rather than a scaling by D (difference of texture and background), we now have a scaling with just "background". It's obvious from the premultiplied blending equation that if background == 0, then A has no affect on the output color.

Obviously when encoding a texture we don't know the value of the background. Looking at a few values :

WPM(A) , background=0   : = 0
WPM(A) , background=1/2 : = 3/4
WPM(A) , background=2/3 : = 4/3
WPM(A) , background=1 :   = 3

it's probably okay to just use WPM(A) = 1, that is just weight all channels the same, use RGBA squared difference. This a compromise given the unknown background; it's convenient that equal weight is not too bad for pre-multiplied alpha. If you have domain-specific knowledge you could use something different. For example on the web or word-processing where you are mostly blending onto white backgrounds, alpha weight closer to 3 would be better.

Conclusion :

When measuring distortion on RGBA textures that you know will be used for alpha-blending, we can find a squared-difference in texture-space that approximates the error in the visible output pixel.

This can be expressed as a weight on the alpha component of the squared difference :

SD(RGBA) = SD(RGB) + W(A) * dA^2

for standard alpha blending :

W(A) = 3 * D^2 / A^2

D = difference between texture and background (scalar)

or W(A) = D_RGB^2 / A^2

for pre-multiplied alpha :

WPM(A) = background_RGB^2

WPM(A) = 3 * background^2

For pre-multiplied alpha, using WPM(A)=1 is reasonable, and just using 4-channel uniform RGBA squared difference is probably fine. This is just another great reason to prefer pre-multiplied alpha.

For non-pre-multiplied alpha, it's hard to escape the need to scale RGB by A for purposes of the error metric. ( W(A) ~ 1/A^2 is equivalent to W(RGB) ~ A^2 which is equivalent to scaling RGB by A). If you aren't scaling RGB by A then you're giving too many bits to RGB when A is low, and not enough when A is high. Assuming you want a scalar weight that doesn't depend on background or A, then plain old uniform W(A) = 1 is not too bad. (D ~ 1/3 and A ~ 1/2 gives W(A) = 4/3 , so W(A)=1 is in the right ballpark).


Requantization to different UNORM bit depths

I wrote before in Topics in Quantization for Games some general basics on quantization. We're going to continue from there, but focus just on the specific case : GPU convention quantization of n-bit unsigned values (UNORM).

As noted before, there are two viable conventions for uniform linear quantizers; either "floor" quantization (with bias on reconstruct) or "round" quantization (bias on quantizer) with direct reconstruction (end points restored directly). The latter is what GPUs use, therefore games should always use this convention for all quantizers to avoid mistakes.

(for example RGBE quantization in the HDR file format uses the opposite convention; floor quantize and bias on dequantize; make sure to use the right convention to match the file format)

The case we will talk about here is :

n-bit value
N = 2^n
quantized values in [0, N-1]
source values in [0.0, 1.0] (inclusive)

quantize(f) = (int)( f * (N-1) + 0.5f )

dequantize(i) = (float) i / (N-1);

quantized 0 -> 0.0
dequantized N-1 -> 1.0

Now we want to talk about "requantization". Requantization is when you have some n-bit value and need to change the representation into an m-bit value; we will assume always that both values are in UNORM GPU convention (bias on quantize).

A correct requantizer is one that corresponds to dequantization and quantization to the new bit depth :

x is quantized in n bits

dequantize :

f = (float) x / (N-1);

quantize to m bits :

y = (int)( f * (M-1) + 0.5f);

y = (int)( x * (float)(M-1)/(N-1) + 0.5f);

y is the requantization of x from n bits to m bits

(N = 2^n , M = 2^m)

Okay so back in Topics in Quantization for Games we showed that "bit replication" is the correct requantizer for changing n to 2n.

In requantization what we focus on is the denominator of the quantizer, so changing n bits to m bits is changing denominator from (N-1) to (M-1). The denominator is also the quantized representation of "One" (1.0 in dequantized range), and we need that to be preserved. The fact that bit replication is the correct requantizer is because bit replication is (N+1) (because N+1 = (1<<n) + 1) and (N+1)*(N-1) (that's (N+1)*One) = (N^2 - 1) which is the new One.

To continue that, we can show that you can repeat bit replication to double bits again :

Start at n bits

Bit replicate once : *= (N+1)

now at 2n bits
2n bits range is N^2

Bit replicate again : *= (N^2+1)

What happens to One in n bits (N-1) :


becomes One in 4n bits

Alternatively just write out the requantizer :

y = (int)( (x/(N-1)) * (N^4-1) + 0.5f );

then using the factorization above this is just :

y = (int)( ((x/(N-1)) *  (N^2-1) * (N^2+1) + 0.5f );
y = (int)( (x * (N+1) * (N^2+1) + 0.5f );

y = x * (N+1) * (N^2+1);

exact in ints , which is bit replicate twice

So for example to go from 4 to 16 bits, you bit-replicate 0xA to 0xAAAA

Note that while bit doubling is the correct requantizer for bit doubling you can *not* just truncate bits to bit half. eg. to requantize from 16 bits to 8 bits, you can not just grab the top 8 bits. While grabbing the top 8 bits would be a round-trip from 8 bit values that had been bit-doubled, it would not put the bucket boundaries in the right place for values that are in between. eg. 16-bit values in [129,255] should requantize to "1" in 8-bit, not just the 0 you would get if you truncated to the top 8 bits.

Okay, so let's get into some requantizers. I want to note that just doing the simple form :

y = (int)( x * (float)(M-1)/(N-1) + 0.5f);
is totally fine and you can just do that and not worry about this. But that said, our goal here is to work out requantizers that are exactly equal to that result, but that stay in integer.

There are a couple general techniques that help with this, so I'll do examples of both.

1. Taylor series in the denominator method.

For example, let's find the requantizer from 16 bit to 8 bit. The simple form is :

y = (int)( x * 255/65535.f + 0.5f);
To find the int form, we can try changing 65535 to 65536. One way to do that is approximate :

1/65535 = 1/(65536-1) = 1/65536 * 1/(1 - 1/65536)

We can use the Taylor expansion of 1/(1 - x) for x small :

1/(1 - x) = 1 + x + O(x^2)

using x = 1/65536 , it's reasonable to drop terms in x^2

So :

1/65535 ~= 1/65536 * ( 1 + 1/65536 )

y = (int)( x * 255/65535.f + 0.5f);
y = (int)( (x * 255 + 32767.5.f)/65535.f );
y = (int)( (x * 255 + 32767.5.f) * (1/65536) * ( 1 + 1/65536 ) );

y = (int)( (x * 255 + 32767.5.f + x*255/65536 + 32767.5.f/65536) * (1/65536) );

65536/255 = 257.003922.. ; try 257
32767.5.f + 32767.5.f/65536 = 32768 - 0.5f/65536 ; try 32768

y ~= (int)( (x * 255 + (x/257.f) + 32768) * (1/65536) );
y ~= (int)( (x * 255 + (x/257.f) + 32768) ) >> 16;

y = (x * 255 + (x/257) + 32768) >> 16;

and in fact that is the correct 16 to 8 requantizer in integers (in the sense that it matches the simple float formula exactly).

There is however an even simpler 16 to 8 requantizer which also matches :

int requantize_16_to_8( int x )
    ASSERT( x >= 0 && x <= 65535 );

    return ( x * 255 + 32768 + 127 )>>16;
which can be made by replacing the (x/257) term with its average.

2. Bit replicating up to make an infinite repeating fraction

Another technique that sometimes works is based on a different conceptual picture.

We have (for example) an 8 bit value; 0x00 is "zero" , and 0xFF is "one" (that's what they dequantize to). To get 0xFF to mean "one" we think of it as a fixed point fraction, and bit replicate so for 0xFF think :

0xFF = 0. FF FF FF ... forever

then there is no number between that and 1.0 , therefore it is equal to 1.0

so to find requantizers or interpolators, we can bit replicate several times and approximate the infinite repetition with a finite number , but then we'll be a little bit too low so we'll have to bias up a bit.

To be clear, this repeated fraction is not just hand having and not just right at the end points; if you could repeat forever it would be exact. We can see it by using our Taylor expansion again and doing all the terms :

1/(1-x) = 1 + x + x^2 + x^3 + x^4 + ...

using x = 1/256 , this is :

256/255 = 1 + 2^-8 + 2^-16 + 2^-24 + 2^-32 + ...


1/255 = shift down 8 + shift down 16 + shift down 24 + ...


0x34/255 = 0 . 34 34 34 34 ...

Any denominator that's a power of two minus one is a repeated binary fraction.

We want to see if we can use just a few repeats and then bias up and get correct requantizer results.

As an example, let's work out the requantizer from 10 bits to 8 bits. As usual the simple way is :

f = x/1023.f
y = (int)( f * 255.f + 0.5f );
but we'll try to find an exact match to that formula that stays in int. We guess that we want to start by making a repeated fraction :
thinking of f as a fraction with decimal 20 places to the left

f = (x << 10) | x;

y = (int)( f * 255.f + 0.5f );

y = ( ((x << 10) | x) * 255 + (1<<19) + bias )>>20;

where we guess we need "bias" because of the fact that we are using a truncated 20 place fraction rather than forever
that is, we need to push up 0.FFFFF slightly to get to 1.0

It turns out it works with bias = (1<<9) (among others)

y = ( x * 1025 * 255 + (1<<19) + (1<<9) )>>20;

( using ( (x << 10) | x ) = x * 1025 )

This is in fact a correct requantizer from 10 to 8 bits :

int requantize_10_to_8( int x )
    ASSERT( x >= 0 && x <= 1023 );

    return ( x * 0x03FCFF + 0x080200 )>>20;
Okay, so we can see it's frequently possible to use an approximation of the infinite repeated fraction to find correct requantizers.

(you could also get this from our one term Taylor expansion as well; essentially we just did 1/1023 ~= 1025/(1024^2) )

This requantize_10_to_8 requires 28-bit intermediates, so it fits in 32-bit registers. For scalar that's fine, but for SIMD we'd like to be able to do these requantizers in as few intermediate bits as possible (ideally in 16 to get wider SIMD).

If we go back to the verbose form we can see how to get these in fewer bits :

y = ( ((x << 10) | x) * 255 + (1<<19) + (1<<9) )>>20;

y = ( ((x*255) << 10) + ( x * 255 ) + (1<<19) + (1<<9) )>>20;

t = (x*255) + (1<<9)

y = ( (t<<10) + t )>>20;

y = ( t + (t>>10) )>>10;

int requantize_10_to_8_alt( int x )
    ASSERT( x >= 0 && x <= 1023 );

    int t = (x*255) + (1<<9);

    return ( t + (t>>10) )>>10;

This gives the same result again; it now uses only 18 bit intermediates.

This form is familiar from the classic "mul8bit" from Jim Blinn's "Three Wrongs Make a Right" :

int mul8bit( int a, int b )
    int t = a*b + 128;
    return (t + (t>>8) )>>8;
mul8bit takes a [0,255] value and makes it act as an interpolator from 0% to 100% , eg. for alpha blending. It is equivalent to :
    (int)( ((a / 255.f) * b) + 0.5f );
which is dequantizing a to UNORM, using it to scale b, then requantizing.

We can now see that the "mul8bit" interpolator is a relative of our requantizers, and we can derive mul8bit by using the repeated fraction or Taylor series.

The end!


This "Blinnish" requantizer is always correct (for to_bits <= fm_bits) :

int requantize_blinnish( int x , int fm_bits, int to_bits )
    int fm_N = (1<<fm_bits) - 1;
    int to_N = (1<<to_bits) - 1;

    ASSERT( x >= 0 && x <= fm_N );

    int t = (x*to_N) + (1<<(fm_bits-1));

    return ( t + (t>>fm_bits) )>>fm_bits;
often you can find simpler requantizers that are correct, but this provides a simple baseline/reference.


Oodle 2.9.0 : old compressors removed

Oodle 2.9.0 is now out. The full changelog is :
## Release 2.9.0 - March 23, 2021

$* *fix* : OodleLZ_Compress had an encoder crash bug in the Optimal level encodes on data in sizes just slightly over a 256KB chunk (eg. 262145) with a repeated substring at the very end

$* *change* : Mac libs and dylibs are now fat binaries with x64 and ARM64
$* *change* : Tex : Oodle Texture no longer checks for license file
$* *change* : defining OODLE_IMPORT_LIB / OODLE_IMPORT_DLL is no longer needed; you can link with either type of lib without setting a define
$* *change* : Oodle public headers no longer define types like U8, SINTa, they are instead OO_U8, OO_SINTa, etc.
$* *change* : Oodle public headers now require stdint.h which on Windows/MSVC means VC2010 or higher

$* *change* : Net : OODLE_PLATFORM_HAS_SELECTDICTIONARYANDTRAIN define removed.  Call OodleNetwork1_SelectDictionarySupported instead.

$* *removed* : Core : support for the deprecated LZ compressors is now removed (LZH,LZA,etc.).  Only newlz (Kraken,Mermaid,Selkie,Leviathan,Hydra) and LZB16 are supported.
$* *removed* : Core : OodleLZ_CompressContext_* functions removed; streaming encoders no longer supported
$* *removed* : Ext : OODLEX_PATH_* public defines removed.
$* *removed* : Ext : OODLEX_WCHAR_SIZE public define removed.
$* *removed* : Tex : OodleTex_CheckLicense func removed ; Oodle Texture no longer checks for license files

$* *deprecation* : OodleConfigValues::m_OodleLZ_Small_Buffer_LZ_Fallback_Size no longer used; newlz compressors no longer ever drop down to LZB16 (default behavior unchanged)
The biggest change is the removal of all deprecated pre-Kraken compressors (LZH,LZHLW,LZNA,LZNIB,BitKnit,etc.) except for LZB16 which stays for the moment (but is also deprecated). Data compressed with the old codecs cannot be decoded by Oodle 2.9.0 ; up to Oodle 2.8.14 data made by any earlier version can always be decoded by later versions.

Note that old versions of the Kraken-family of compressors still be decoded (Oodle 2.9.0 will read data written by Oodle 2.3 Kraken), and that will always be the case, we never break reading old data (made by the supported codecs). So if you are using the Kraken-family of compressors, you can always update to newer versions and your old data will be readable.

If you were using one of the old codecs, we recommend changing to one of the Oceanic Cryptozoology codecs (Kraken, Mermaid, Selkie, Leviathan). They are better in every way. To do this, you need to decode that data with the old version of Oodle you have (or any version up through 2.8.14) because 2.9.0 cannot decode the old codecs. Assuming your old version is post-Kraken (2.1.5), you can re-encode to Kraken with your old sdk and the current version (2.9.0) will be able to load that data.

If you can't switch codecs for some reason, you can always keep using the old version of Oodle you are currently on.

We have also designated the last version before 2.9.0 (which was 2.8.14) as a "long term support" release. We will continue to push critical bug fixes to 2.8.14.lts for people who need to stay on old codecs.

(2.8.14.lts has already received a bug fix since the 2.8.14 first release; clients on 2.8.14 should update to the latest version)

If at all possible we recommend everyone to get on the Oceanic Cryptozoology codecs and update to 2.9.0 ; if you need the old codecs we recommend 2.8.14 for long term support.

LZB16 is still supported in 2.9.0 because some data may need it, but we strongly discourage its use.


Faster Inverse BWT

The BWT (Burrows Wheeler Transform) has long fascinated people for its ability to capture complex correlations with a very simple inverse transform. Unfortunately despite that inverse transform being very simple, it is also slow. I will briefly review the inverse BWT (i-BWT) and then look at ways to speed it up.

Jump to the end for the punch line : speed results and source code.

Let's briefly review the basic BWT to establish the baseline algorithm and notation.

Given an input string like "inputstring" add an EOF symbol so S = "inputstring|" then form all rotations :

Then sort all the rows, this gives the matrix M :
^          ^
F          L
this is all the suffixes of the file, and we've sorted them all. We've labeled the first column F and the last column L.

The last column L is the BWT, which is what we transmit. (typically we don't transmit the EOF character, instead we transmit its location and leave it out of the BWT string that is sent).

In the matrix M each column is a permutation of the original string S (remember the rows are cyclic rotations). The first column F is just the alphabetic sort of S. F can obviously be reconstructed from a histogram of S, which can be counted from L. The column L is the original characters of S in order of their following suffixes.

The inverse transform is simple conceptually. We transmit L, the BWT string (the last column), we need to regenerate S the top row.

Because the rows are all just rotations of S, the character at F[i] on each row is the character that follows L[i]. This gives you an answer to "what's the next character after this one?" which is all you need :

Output decoded string D
Start with D=| the EOF character.

What's the next character?

L = nr|iinttsupg
F = giinnprsttu|

Find | in L , it's L[2]
so the next char is F[2] = i


What's the next character?

Find i in L , it's L[4]  (NOT L[3])
so the next char is F[4] = n

Then find n at L[5] (not L[0])
the next char is p, so D=|inp

after p is u, etc.
So you can just read out the characters one by one like that.

When a character occured multiple times in L we had to find the right one (which corresponds to the same location of that character in the original string S). eg. when finding 'i' we needed L[4] not L[3] because L[4] is the i that follows | in "input", while L[3] is the i in "string".

The way to do that is to find the same occurance number of that character. If it's the 2nd "i" in F, we want that to map to the 2nd "i" in L.

The reason is because L has the characters sorted by their suffixes, while F is sorted by the whole strings. So when they are the same symbol in L, then the order they occur in L is the same as the order the occur in F.

For example the n's in F :

are in this order because "ng" is less than "np"; in the last column L, the n's occur as :
they are sorted by their suffixes "g|" and "pu", so they are in the same order.

So the nth occurance of each character in L/F corresponds to the nth occurance of each character in L/F. The main thing we need for inversion is this mapping between L and F for each character.

It turns out to be more convenient to go backwards and make the mapping from L to F (rather than F to L). This is because the F array can just be implicit, we never need to make it in memory. F is just the "cum prob" table. That is :

Each symbol has a histogram count[s]
start[s] = prefix sum of counts < s
cumulative probability range[s] = [ start[s] , start[s] + count[s] )
(closed interval on low end, open interval on high end)
F[] = fill s in range[s]

To find the nth occurance of s in F[] is just i = start[s] + n
This is the same array you would use in arithmetic coding or RANS.

We'll calling the mapping from L to F , LF[] and it is just constructed thusly :

scan L[]
s = L[i]
count C[s] as you go (start at zero)
find the C[s]th occurance of s in F[]
LF[i] = start[s] + C[s]
then C[s] ++
Now LF[i] takes us from the symbol at L[i] to the same symbol in F[] which will be at F[ LF[i] ], then from there you can step to the prior symbol, which is at L[ LF[i] ].

Note that this is backwards from what we were doing before, so we are now reading at the stream in reverse. To read out the i-BWT (backwards) is now : :

start at i = location of EOF in L[]

for N times :
find symbol L[i] in F[] ; this is i = LF[i]
then step to the prior symbol F -> L
output L[i]
that's it! A full BWT inverse transform is then :
L[] = the BWT string, we get this from the decoder

count[i] = histogram of L[]
start[i] = prefix sum of count[]

make LF :

C[] = 0
for i to N
    s = L[i];
    LF[i] = start[s] + C[s];

read out :
i = index of EOF (transmitted)
// L[i] is EOF sym which we don't need to output
for N times
    i = LF[i];
    *ptr++ = L[i];
This produces the string backwards, which we usually handle by reversing in the encoder before doing the BWT so that the decoder can just output forwards.

That's it!

I'm not going to talk about how the compressor transmits the BWT string L[] compactly. See the original BW94 paper or the bzip code for the classic MTF + RLE + switching Huffmans encoding. See Fenwick for analysis of BWT as a symbol ranker. In the modern world most people do BWT encoding without MTF; see Maniscalco M99, context induction, BCM and BBB.

I'm also not talking about the forward transform at all. The forward transform also has interesting efficiency issues, mainly also relating to cache misses for large buffers or to GPU implementations, and there has been lots of good research on that; see the papers. In this blog I will assume you don't care much about forward transform speed, so I'll push all the slow operations to the encoder side to make the decoder as fast as possible.

I will also just briefly mention that BWT of long string recurrances can be sped up with an LZP pre-pass. We will assume this has been done where appropriate so that there are no very long identical strings in our source input (eg. perhaps suffixes are always unique after 16 symbols).

End of review and we'll dive into speeding up the i-BWT.

So the transform is delightfully simple, but it's slow. Why?

(and by "slow" I mean *really* slow; specifically 200-400 cycles per iteration on my Intel machine "beasty")

The problem is that each instruction of the i-BWT intrinsically depends on the previous instruction (instrinsically meaning can't be removed with code reorganization), so no instruction-level-parallelism (ILP) can be used. On modern CPUs that might be able to do 4 instructions per clock this is bad. What's worse is that there is an intrinsic cache miss in the lookup step of the i-BWT.

The problem is the index of the next symbol in the sequence is semi-random and goes all over the BWT buffer. It cache misses often (almost every single iteration). Now you can obviously fix this by making the BWT chunk smaller, which is what the old bzip does. If the chunk is small enough to fit in cache, no problem. But that also kills compression and sort of ruins the whole point of the BWT (if small chunk BWT compression is good enough for you, then you should just use an LZ like Kraken which better fits that space-speed niche; high speed small cache is outside the operating range of where BWT makes sense vs alternatives). In-cache BWT is not a Pareto optimal way to do a fast compressor; we are only interested in BWT in the high-compression domain. So we want to use BWT chunk sizes that are bigger than the fast cache (16M chunks perhaps, which often fits in L3) and speed things up.

The core of the i-BWT is just this :

for N times
    i = LF[i];
    *ptr++ = L[i];
The problem is the LF[i] are a near-random walk around the buffer. They are in order of the sorted suffixes, which tend to be completely randomly distributed in N, assuming the input array is not in some degenerate pre-sorted order. You load at LF[i] and it cache misses, you can't do any more work, the loop completely stalls. As soon as the cache miss returns, you immediately do i = LF[i] and look up at LF[i] again and stall again.

Now one obvious thing we can do is to first combine the L[] and LF[] arrays into a single merged array, so that we get exactly 1 cache missing load per iteration instead of 2. But we still need fundamental improvements.

There are two big deficiencies we need to address :

1. Instruction sequence with sequential dependencies
2. Cache miss on looking up the next symbol

The most obvious approach to deal with these deficiencies is to use a technique we are now familiar with, which is to do multiple streams at the same time.

Independent multiple streams allow you to keep more execution units busy, especially in a case like this where there is a fundamental cache miss stall that you want to fill with useful CPU work during those cycles.

Say you have a single execution stream where every instruction is dependent on the last (c follows b follows a), only 1 can execute per cycle, and if one cache misses you get unused cycles with no work to do :

clock tick on the left
"1" is the independent stream number
a,b,c indicate a sequence of dependent instructions

0: 1a
1: 1b
2: 1c (cache miss)
3: .
4: .
5: .
6: .
7: 1d

If you run 4 streams of independent work, each stream dependent on the last instruction within its stream but no cross-stream dependencies, now the CPU can execute 2 instructions per cycle (say), and fill in during the cache miss :

0: 1a + 2a
1: 1b + 2b
2: 1c + 2c (both cache miss)
3: 3a + 4a
4: 3b + 4b
5: 3c + 4c (both cache miss)
6: .
7: 1d + 2d

It's a bit like what hyper-threading does for you, but we get it just by running more independent execution sequences on our single thread to give the out-of-order execution a bigger selection of work to find instructions to fill the gaps.

In our case with heavy cache missing dependent work, in the 1-stream case you are waiting on the latency of each cache miss with the processor doing nothing, so you get a full stall all that time. Going to more streams at least lets us get multiple memory operations pending at once, rather than stall, wake up, stall again on one load at a time.

In the case of i-BWT we can do this by taking the input data (before BWT) and think of it as T segments. We don't cut the string into pieces (remember chunking hurts compression and we don't want to do that), we still do the forward BWT on the whole string.

Like :


in 2 segments :

do the BWT on the whole string "inputstring|"

then find the locations of the EOF symbol | in L[]
and the + segment boundary, by finding where "string|" starts

inputstring| <- EOF key
string|input <- segment key
^          ^
F          L

EOF key = 2
segment key = 7

Then you would transmit the BWT string L[] and the EOF key (2) as usual, and also transmit the segment key (7). The segment key lets you start the i-BWT core loop at that location and read out characters from there, in addition to the stream from the EOF key.

In the i-BWT you can start output cursors for all T segments and you can read them out of the i-BWT simultaneously. The i-BWT core loop just does "from this symbol, what's the next symbol?" so you can do T of those independently.


cursor 1 starts at i=2 in L[]
cursor 2 starts at i=7 in L[]


The N-stream decode is very simple, literally just doing the 1-stream process to N output pointers :

T segments
key[T] are the start indexes
i1,i2,etc. = key[]
ptr1 = start of output buffer
ptr2 = ptr1 + (N/T) , start of next segment, etc.

for N/T times
    i1 = LF[i1];
    *ptr1++ = L[i1];

    i2 = LF[i2];
    *ptr2++ = L[i2];

it just does the basic i-BWT loop on T segments in each iteration. Crucially these are independent, so when one of them stalls on a cache miss, the others can still do work. If all of them were stalling on cache misses and the memory system could service T cache line fetches simultaneously, we would be able to stall with all T requests in flight which would give us factor of T speedup. In practice it appears to be somewhat less than that.

You can also cut into segments like this for parallel decoding on multiple threads for very large blocks (without doing any chunking, which sacrifices compression). eg. you might find 16 segments to decode and give 4 segments each to 4 threads. The threads have no cache contention because they are writing linearly to different output segments, and the shared memory of the LF[] array is read-only in this loop.

Going to 4 streams provides around a 2.7X speedup, 8 streams gets us about 4X :

enwik7 :
ibwt_1x1 : seconds:0.7500 | cycles per: 299.330 | MB/s : 13.33
ibwt_1x4 : seconds:0.2801 | cycles per: 111.769 | MB/s : 35.71
ibwt_1x8 : seconds:0.1835 | cycles per: 73.218 | MB/s : 54.51
(see more speeds at end)

The next thing we can do is an idea from the paper "Cache Friendly Burrows-Wheeler Inversion". We can make our basic i-BWT step output words or dwords instead of bytes.

This is a simple idea, but I think it's quite interesting *why* and *when* it works.

What we want to do is just do our core i-BWT to output words at a time instead of bytes, by outputting a precomputed pair of one byte steps :

core i-BWT loop for words :

i = key (EOF location in L[])
for N/2 times
    *word_ptr++ = LL[i];
    i = LF2[i];
where LL = a word, two steps of L[] , and LF2[] = two steps of LF[], that is LF2[i] = LF[LF[i]].

(reminder that in practice we put the LL[] and LF2[] into a single struct so that we get one cache miss array lookup rather than two).

So to use that we just need to build the LL and LF2 arrays first. To do that we just need to precompute each pair of steps in the byte-at-a-time BWT :

prep for word i-BWT :

first build LF[] as usual

for i = 0 to N
    int i2 = LF[i];
    LL[i] = L[i] | (L[i2]<<8);
    LF2[i] = LF[i2];
At every index in L[] we precompute a two-character step by following LF[] once.

(side note : you could precompute a two-character step forward more simply; from L[i] the next character is just F[i], which you could get from the cumulative probability table at i, but that would give you a step forward and what we want is a step backward; L[i2] here is the character that precedes L[i]; it wouldn't help because we need a lookup at i2 to make LF2[] anyway)

Now at first glance this might not seem helpful, what we have done is transform :

byte at-a-time BWT :

core i-BWT loop :

N times :
    i <- LF[i]
    output byte from i

word at-a-time BWT :

prep :

N times : 
    LF2[i] = LF[LF[i]]
    word = L[i] + L[LF[i]]

N/2 times :
    i <- LF2[i]
    output word from i

We are actually doing more work; we now do 3/2 N iterations instead of N iterations. But it is in fact much faster.


The reason is that the N-step "prep" loop to build LF2 does not have the two big problems of the core loop. Remember the two big issues :

1. Instruction sequence with sequential dependencies
2. Cache miss on looking up the next symbol
these are what make the "core" i-BWT loop so slow.

First issue #1 : the "prep" loop naively allows for execution ahead, unlike the core loop. The reason is that it is just doing i++ to iterate the loop, rather than chasing i = LF[i] around a data-dependent serpentine walk through the arrays. This means that the processor can effectively unroll the loop to execute ahead :

prep loop :

iter 1 :
    LF2[i] = LF[LF[i]]
    word = L[i] + L[LF[i]]

iter 2 :
    LF2[i] = LF[LF[i]]
    word = L[i] + L[LF[i]]
iter 1 and iter 2 can run at the same time because i++ can be precomputed.
If LF[LF[i]] stalls on a cache miss in iter 1, iter 2 can still proceed.

core loop :

iter 1 :
    i = LF[i]
    output byte/word from i

iter 2 :
    i = LF[i]
    output byte/word from i

iter 2 can't start until iter 1 is done because we don't know "i" until the
results of the lookup at LF[i] are done.
A modern processor will be able to execute ahead in the "prep" loop to fully saturate execution, and we don't need to manually do the N-streams to get ILP because it just naturally exists.

issue #2: cache misses. The "prep" loop cache misses far less than the "core" loop.

Again this is not naively obvious. The byte "core" loop does N lookups at LF[i]. The "prep" loop also does N lookups at LF[i]. So they should cache miss the same, right?

Nope. The difference is because the "core" loop is doing i = LF[i] which takes you to a semi-random place each time, whereas the "prep" loop is doing i++ for subsequent lookups, and that means the lookups are often to nearby locations that stay in cache.

core loop does :

    starting from i
    i1 = LF[i]
    lookup at i1
    i2 = LF[i1]
    lookup at i2
    i3 = LF[i2]
    lookup at i3

prep loop does :

    starting from i
    i1 = LF[i]
    lookup at i1
    i2 = LF[i+1]
    lookup at i2
    i3 = LF[i+2]
    lookup at i3
The indexes i1,i2,i3 in the core loop are semi-random, in the prep loop they are not.

To understand, let's remember what LF[] is. It takes us from the symbol at L[i] to the location of the preceding symbol in the BWT (we're outputting backwards).

So say you were currently at "ing|" (in "inputstring|"), it takes you to the location of "ring|" then "tring|" ; those are i1,i2,i3 in the core loop.

In the "prep" loop you are taking the starting strings at i, i+1, i+2. These are subsequent in sorted order. These might be something like :

"hella ..."
"hellish a.."
"hellish b.."
"hello a..."
"hello b..."
So the steps LF[] gives us to i1,i2,i3 will be to tack on the preceding character and look that up. They might all be preceded by random characters, but in practice because these strings are similar their preceding characters tend to be similar. That might be :
" hella ..."
" hellish a.."
"shellish b.."
" hello a..."
"shello b..."
then you look up those locations in the suffix sort. As long as you tacked on the same preceding character, they will stay adjacent :
" hella ..." -> i =4000
" hellish a.." -> i = 4001
"shellish b.." -> i = 6078
" hello a..." -> i = 4002
"shello b..." -> i = 6079
The preceding character for the sorted suffixes is of course just the BWT string L[] itself. The portion here would be " s s".

So as long as the BWT string has repeated characters, the indexes of the "prep" loop lookup will be adjacent. If they are adjacent indexes, they will not cache miss because we already loaded their neighbor which brought in that cache line (load at i=4001 won't cache miss because we just did a load at i=4000).

Well, getting repeated characters in L[] is exactly what the BWT gives us! If we think of it in terms of post-MTF BWT strings, a 0 will be a load from the most likely location, 1 will be the second most likely, etc. so there will tend to be a few cache lines that we already have that provide most of our loads for the "prep" loop. Only rarely do you get unexpected characters that cause cache misses.

Note that this only works on *compressible* data where the BWT is giving the coherence we want. On random data that doesn't compress this breaks down. (though even in that case, there are still at most 256 hot cache lines = 16 KB that we need to read from, so it's still better than the "core" loop, and we still have property #1 that each iteration is independent).

Actual load locations so you can see how this works in practice :

LF[i] = 8853
LF[i] = 8854
LF[i] = 8855
LF[i] = 8856
LF[i] = 10553
LF[i] = 8857
LF[i] = 10554
LF[i] = 48
LF[i] = 49
LF[i] = 50
LF[i] = 10555
LF[i] = 10556
LF[i] = 51
LF[i] = 52
LF[i] = 5070
LF[i] = 10557
LF[i] = 2477
LF[i] = 53
LF[i] = 54
Where you have 10553,10554,10555 that means the same character was preceding those suffixes, so they lookup in adjacent places.

You can of course take this idea for word output and repeat it again to get dword (u32) output per iteration. In practice what I see is that dword output is only faster for compressible data where the cache coherence property above works well. On random input dword output gets a lot slower than word-at-a-time. Because of this I think word output is best if you only choose one i-BWT, but you can do even better by adaptively choosing the output algorithm by the compression ratio.

The dword (4x) variant does an N-loop to build LF2 which is quite coherent, the next N-loop to build LF4 is slightly less coherent unless the data is compressible (we're relying on two-symbols-away correlation now instead of neighbor correlation, which is less strong), then it does N/4 of the "core" loop which is still quite cache missy. So the net is :

1x byte loop : N steps of "core" , cache mising
2x word loop : N steps of LF2 + (N/2) steps of core
4x word loop : N steps of LF2 + N steps of LF4 + (N/4) steps of core

The authors of "Cache Friendly Burrows-Wheeler Inversion" have another method that accelerates long repeated substrings called "precopy". I think that a pre-pass with an LZP transform is probably a faster to way to accomplish the same thing on data where that is helpful. "precopy" also makes N-stream interleaving much more complex. I have not tested it.

Speeds measured on my Intel Core i7 3770 (locked at 3403 MHz) (Ivy Bridge)

ibwt 1x1 = classic byte at a time inverse BWT
ibwt 1x4 = byte x 4 streams
ibwt 2x4 = word x 4 streams
ibwt 4x4 = dword x 4 streams
ibwt 1x8 = byte x 8 streams
ibwt 2x8 = word x 8 streams
ibwt 4x8 = dword x 8 streams

cycles per byte for the ibwt :

name 1x1 1x4 2x4 4x4
dickens 350.6 123.7 70.0 47.9
enwik7 301.2 116.0 66.8 47.0
webster 376.9 139.0 78.5 51.2
lzt99 211.3 114.6 65.1 61.4
rand_16M 401.6 130.6 80.4 297.3

name 1x1 1x8 2x8 4x8
dickens 337.0 79.8 46.8 35.8
enwik7 299.1 73.2 46.0 36.1
webster 376.5 98.0 57.5 40.3
lzt99 208.6 77.9 48.1 53.7
rand_16M 401.3 84.1 55.0 273.4

We've improved the i-BWT speed from around 300 cycles/byte to 50 cycles/byte. That's still extremely slow. It's around the speed of Oodle LZNA (30 cycles/bytes) or 7z LZMA (70 cycles/byte). It's an order of magnitude slower than Oodle Leviathan (5 cycles/byte). But in some cases, such as text, BWT on large blocks gets a lot more compression than those, so it could be interesting on the right data types if you care about the very high compression domain. All speeds single threaded.

The i-BWT on a small buffer that stays in cache is around 10 cycles/byte (and could probably be faster; we haven't looked at all at micro-optimizing the actual instructions here, since if we're stalling on cache misses they don't matter), so at 50 cycles/byte we're still spending 40 cycles/byte stalled on the memory subsystem in our best algorithm, which is not great.


name 1x1 1x8 2x8 4x8
dickens 301.3 52.4 33.9 27.7
enwik7 265.8 50.5 33.1 28.2
webster 319.6 54.5 34.8 29.0
lzt99 177.5 50.6 32.5 35.6
rand_16M 336.6 53.5 41.4 154.1

Without large pages, the bottleneck appears to be in TLB/page mapping. With large pages it seems to be limited by the rate that cache missing reads can be retired (assuming the latency of that is fixed, this is the number of different cache lines that can have in-flight reads simultaneously). eg. if latency is 300 cycles, with 10 simultaneous reads we could get to 30 cycles/byte.

Downloads :


fast_ibwt_test exe (Win x64)

The source code is public domain. The ibwt routines are standard C. The test driver uses cblib


Rate allocation in Oodle Texture

Oodle Texture does rate-distortion optimization (RDO) BCN encoding, optimizing the BCN encoding for an R-D two axis score such that the size of the BCN texture after a following lossless compression is reduced while the distortion (difference from original) is not increased too much.

One way to think about RDO conceptually is as rate allocation. Rate allocation is when an encoder intentionally puts more or fewer bits in parts of the data to control what the decoder will receive, so that it can use more bits where it helps more, and the result is a better quality for a given output size. I want to talk a bit about that viewpoint and how it works in Oodle Texture.

Traditional lossy encoders like JPEG didn't have active rate allocation; they used a fixed scheme (DCT and quantize) that did a fixed rate allocation (give more bits to the lower frequencies) and didn't try to optimize that based on image contents. Modern codecs like H264 use rate allocating encoders and furthermore have explicit tools in the format to make rate allocation more flexible, such as variable quantizers. Formats without explicit rate allocation controls can still be rate adjusted, which is what we do in Oodle Texture, but it's not as easy to dial.

In that context, we think of a traditional non-RDO BCN encoder as allocating rate evenly to all blocks; it doesn't care where the rate goes and just tries to minimize distortion. The BCN encoder itself produces fixed size blocks (8 bytes per block for BC1 or 16 bytes per block for BC7), but the "rate" we care about is the size of the block after subsequent LZ compression (eg. with Kraken or Zip/deflate).

In Oodle Texture, we use the Lagrange parameter method for RDO. Specifically we construct a Lagrange cost :

J = D + lambda * R
and then we can just minimize a single number, J, rather than a two-axis score of {R,D}. Obviously when lambda=0 this reduces to J=D , minimize D, which is a non-RDO encoding that only cares about quality and not rate. As lambda increases your score is more and more a balance of minimizing both distortion and rate.

We sometimes think of D(R) , distortion just being a function of R, but that's not how it works in an implementation, you don't actually know the functions for how R and D relate.

In practice there is some encoding, which I will index by "i" , and you have two separate functions :

R(i) and D(i)
That is, you just choose various encodings that look promising, and measure R and D of each. This will give you a scattering of points on an R-D plot. Some points are off the Pareto Frontier - strictly worse in terms of R & D and we just reject those points. See Fabian's blog on RDO for some pictures of this.

Henceforth I will assume that points off the Pareto Frontier have been rejected and we now only are looking at encodings on the Pareto R-D curve. Also I will sort the index of encodings "i" by rate.

The R(i) and D(i) curves both tend to be hyperbolic-shaped, with minima at opposite ends of the "i" spectrum. If you minimized either one on their own, they don't have any local minima so you would just go to the edge of what the encoding format is capable of doing. By adding them together with lambda, it makes a trough somewhere in the middle :

High D, low R on the left (very lossy), low D high R (non-RDO) on the right.

At low lambda, the minimum of J is in the high-R, low-D domain. As lambda increases it shifts the minimum of J to lower R. In the example chart, the red low-lambda curve has minimum at i=14, the teal medium-lambda curve has minimum at i=12, the gold high-lambda curve has minimum at i=10. In this way, lambda can be used to dial rate, but indirectly.

Concretely, if we reparameterize to imagine we have a function D(R) (Distortion as a function of R; there is only one encoding for each R because we discard non-Pareto encodings), we have :

J(R) = D(R) + lambda * R

minimum is at d/dR J = 0

d/dR J = 0 = d D /dR + lambda

lambda = - dD / dR

lambda is the slope on the D(R) curve.

Because the D(R) curve is hyperbolic-shaped, the slope acts like a parameter of D itself. That is, where D is higher, the curve is also steeper, so by dialing the slope we have control over the principle value D as well.

Aside for experts : we are assuming that D(R) is continuous and monotonic AND dD/dR is continuous and monotonic; that is, the slope steadily increases from low to high. The whole idea of lambda as a slope of D(R) only really works if these curves are smooth and hyperbolic shaped as expected. We also need J to only have one local minimum. If these assumptions fail (which in practice they do!) you can wind up at a minimum that's not where you want, and the lambda-dials-D relationship can break down. There are some tricks to avoid these pitfalls. Try to avoid large steps of R/D in your encoding choices; provide the encoder with a fine grained series of steps; don't use the true R which can have strange non-monotonic wiggles, instead use an approximation of R which tends to smooth things out; you generally want to pre-classify the encodings that you think are low R vs high R and force them to be monotonic rather than measuring true R.

Now, "lambda is the slope on the D(R) curve" sounds a bit unintuitive and abstract, but in fact it is a concrete simple thing.

Lambda is a price. It is the gain in D that you must get for a gain in R to be worth it.

By using lambda as your control parameter instead of D, you are dialing quality via the *cost* of quality in terms of rate. You are saying "this much gain in quality is worth this much rate to me". We typically measure R in bits, so let's say lambda is quality per bit. In order to make a file bigger by 1 bit, it must provide at least "lambda" quality gain. Smaller quality gain, it's not worth it, don't do it.

Having our control paramter be price instead of directly controlling R or D turns out to be very beneficial, both for us internally and for you externally.

First why it's right externally : lambda control (slope control) lets you set a single parameter that works for all kinds of data or images. It automatically gives more rate to images that are harder to compress ("harder" here very concretely means a steeper slope - they give up more distortion than average for each step or rate). So you can have a mix of very vastly different data, lightmaps and normal maps and cartoon drawings, and they automatically rate adjust to send bits where they help the most. Many novices prefer to think of "rate reduction" or a "constant quality" kind of setting, like "I want to RDO all my images to 30% rate reduction", but that's really wrong. On some images, getting to 30% rate reduction would do lots of quality damage, while on others you could easily get more rate reduction (say 50%) without much quality loss at all. Lambda does that for you automatically.

Internally it's important because it lets us make each coding decision in isolation, just by looking at the J cost of that one decision in isolation. It makes the problem separable (which is also great for parallelism), but still achieve a global minimum. Lagrange optimization automatically does rate allocation within the image to different blocks and decisions.

I think it helps to compare to an alternative manual rate allocation method to see how advantageous it is :

Algorithm Manual Rate Allocation :

you have an image of N BCN 4x4 blocks
for each block, find the lowest D (non-RDO) encoding
measure R of that encoding - this is the max rate for each block

now incrementally reduce total rate
you want to take rate from the best block to take rate from of the N

for all N blocks
find the next smaller encoding 
measure the D increase for that step down of R
slope(n) = delta(D) / delta(R)

take the one block change n with the lowest slope(n)

repeat until all slopes(n) are > max_desired_slope


Algorithm Lagrange Rate Allocation :

for all N blocks in parallel :

try various encodings of the block
compute J = D + lambda * R for each encoding
take the encoding on that block with the best J

In "Manual Rate Allocation" we have this palette of various bins (the blocks) to choose to take rate from, and you take it from the one that does the least damage to D, if you keep repeating that the slopes will converge until they are all roughly equal.

The powerful thing is that these two algorithms converge to exactly the same solution (caveat: assuming monotonic and smooth R/D curves, which in fact they are not). The Lagrange method is actually doing rate allocation between the blocks, even though it's not explicit, and each block can be considered in isolation. In an RDO encoder, rate is like water, it flows away from high ground (dD/dR is high) to low ground, until level is reached. The Lagrange method is able to do this separably because the total amount of water (rate) is not conserved.

We can control exactly what the rate allocation does through the choice of the D function. The R function for rate is relatively inflexible - we're counting bits of output size (though in practice we may want to use a smoothed rate rather than true rate), we don't have much choice in R - in contrast, the D function is not absolutely specified and there are lots of options there. The choice of D changes where rate flows.

For example consider the common choices of D = L1 norm (SAD) vs L2 norm (SSD). These metric score different errors differently, which causes rate to flow. In particular L2 (squaring) puts a very big weight on single large errors vs. smaller multiple errors.

Source data   = { 10, 10 }
Approximation = {  8, 12 }

SAD error = 2 + 2 = 4
SSD error = 2^2 + 2^2 = 8

Source data   = { 10, 10 }
Approximation = { 10, 14 }

SAD error = 0 + 4 = 4
SSD error = 0^2 + 4^2 = 16

Choosing D = SSD would cause rate to flow from the {8,12} approximation to the {10,14} approximation because the error is much bigger there (relative to where it is with SAD). All reasonable D functions will be zero for an exact match, and increase as the approximation gets further from the source data, but they can increase at different rates for different types of distortions. In a non-RDO encoding, these different D functions might find very similar encodings, but with RDO by choosing different D you get rate to flow between blocks to different characters of error.

With no further ado, we can get to the fun part : pictures of how rate allocation plays out in practice in Oodle Texture.

These images show the original image, followed by gray scale images of the rate allocation. These are for BC7 encodings with Oodle Texture. The non-RDO encoding is Oodle Texture at lambda=0 ; the RDO encodings are at lambda=40 (typical medium quality setting).

In the rate allocation images, each pixel represents one 4x4 block. White is full rate (16 bytes per block) and black is zero bits. These encodings are not normalized to the same total rate (in particular the non-RDO is of course much larger). The images are gamma corrected so that linear light intensity corresponds to bits per block. The original images are halved in size for display here. So for example "mysoup" was 1024x512, the original shown here is 512x256, and rate map is one pixel per block, so 256x128.

The "rmse" and "perceptual" images use the exact same coding procedure at the same lambda, they differ only in the D function used to measure distortion. "rmse" tries to minimize simple L2 norm, while "perceptual" has some estimation of visual quality (trying to avoid blocking artifacts, for example).

The R measured for these maps is the size of each block after subsequent compression by Oodle's LZA compressor.

non-RDORDO rmseRDO perceptual

The non-RDO "mysoup" has very high entropy, the BC7 blocks are nearly incompressible by the LZ (7.759 bpb). The RDO with D=rmse keeps very high bit rate in the blocks of the bowl rim, but allocates lots of bits away from the smooth regions in the egg and pumpkin. In the "perceptual" allocation, the bits in the rim are reduced, allowing more error there, and the bit rate of the smooth regions come up to avoid blocking artifacts.

non-RDORDO rmseRDO perceptual

In the house image "test6" the bit rate of the "rmse" allocation goes nearly to zero in the smooth dark areas of the beams under the eaves. That's a mistake perceptually and causes bad blocking artifacts, so the "perceptual" allocator shifts rate back to those blocks.

non-RDORDO rmseRDO perceptual

We can see that Oodle Texture is changing near-even rate allocation to very variable rate per block. Even though the BC7 blocks are always 16 bytes, we have made some more compressible than others, shifting rate to where it is needed. On many blocks, 16 bytes is just too much, that much rate is not needed to get a good encoding, and the difference in D to a reduced-size encoding is small; these are low slope blocks and will lose rate in RDO. After rate allocation, the blocks all have the same dD/dR slope; they reach an equilibrium where you can't take bits from one block and move it to another block and improve total quality.

Something you may notice in all the rate allocation maps is there are horizontal lines of higher rate going through the image. This is where we slice the images into chunks for parallelism of some operations that have to work per-chunk instead of per-block. The stripes of higher rate show that we could find some improvement there.

Links :

Oodle Texture at radgametools.com
Lagrange space-speed optimization - cbloom blog
Leviathan is a market trader - cbloom blog
Rate-Distortion Optimisation in Dirac
Rate-distortion optimization The ryg blog


Oodle 2.8.14 with Mac ARM64

Oodle 2.8.14 is out. The full changelog is at RAD. The highlights are :
## Release 2.8.14 - February 15, 2021

$* *enhancement* : BC7 encoding is faster ; slightly different encodings at higher speed with similar quality

$* *new* : Mac ARM64 build now provided ; Mac example exes are fat x64+arm64
$* *new* : Apple tvOS build now provided

$* *deprecation* : Mac 32 bit x86 build no longer provided
We're also now shipping plugin integrations for Unreal 4.26

Kraken decompression is wicked fast on the M1 :

Kraken, Win-x64 msvc-1916, lzc99.kraken.zl6
lzt99 : 24,700,820 ->10,013,788 =  3.243 bpb =  2.467 to 1
decode           : 16.444 millis, 2.26 c/B, rate= 1502.09 MB/s

Kraken, Mac-x64 xcode-12.0.0, lzc99.kraken.zl6
lzt99 : 24,700,820 ->10,013,788 =  3.243 bpb =  2.467 to 1
decode           : 16.183 millis, 2.10 c/B, rate= 1526.36 MB/s

Kraken, Mac-ARM64 xcode-12.0.0, lzc99.kraken.zl6
lzt99 : 24,700,820 ->10,013,788 =  3.243 bpb =  2.467 to 1
decode           : 11.967 millis, 1.55 c/B, rate= 2064.13 MB/s

Win64 run is :
AMD Ryzen 9 3950X (CPU locked at 3393 MHz, no turbo)
Zen2, 16 cores (32 hyper), TSC at 3490 MHz 

Mac runs on Macbook Mini M1 at 3205 MHz

Mac x64 is emulated on the M1
c/B = cycles per byte should be taken with some salt as we have trouble finding real clock rates, but it's clear the M1 has superior IPC (instructions per clock) to the Zen2. It seems to be about the same speed as the Zen2 in emulated x64!

It will be interesting to see what the M1 high performance variants can do.

Some more speeds cuz I like big numbers :

Macbook Mini M1 ARM64 :

Mermaid, Normal, lzt99 :
24,700,820 ->11,189,930 =  3.624 bpb =  2.207 to 1
encode (x1)      : 299.154 millis, 37.21 c/B, rate= 82.57 MB/s
decode (x30)     : 6.438 millis, 0.80 c/B, rate= 3836.60 MB/s

Mermaid, Optimal2, lzt99 :
24,700,820 ->10,381,175 =  3.362 bpb =  2.379 to 1
encode (x1)      : 3.292 seconds, 409.43 c/B, rate= 7.50 MB/s
decode (x30)     : 7.134 millis, 0.89 c/B, rate= 3462.57 MB/s

Selkie, Normal, lzt99 :
24,700,820 ->13,258,742 =  4.294 bpb =  1.863 to 1
encode (x1)      : 213.197 millis, 26.51 c/B, rate= 115.86 MB/s
decode (x30)     : 3.126 millis, 0.39 c/B, rate= 7901.52 MB/s

Selkie, Optimal2, lzt99 :
24,700,820 ->12,712,659 =  4.117 bpb =  1.943 to 1
encode (x1)      : 1.861 seconds, 231.49 c/B, rate= 13.27 MB/s
decode (x30)     : 3.102 millis, 0.39 c/B, rate= 7962.55 MB/s



AVIF is an image format derived from I-frames of AV1 video (similar to HEIC/HEIF from H265/HEVC). See also my 2014 Test of BPG , which is an H265 I-frame image format.

Here are some links I've found on AVIF :

AVIF image format supported by Cloudflare Image Resizing
GitHub - AOMediaCodeclibavif libavif - Library for encoding and decoding .avif files
GitHub - googlebrunsli Practical JPEG Repacker
Releases · kornelskicavif-rs · GitHub
GitHub - link-ucavif avif encoder, using libaom directly.
GitHub - xiphrav1e The fastest and safest AV1 encoder.
AVIF for Next-Generation Image Coding by Netflix Technology Blog Netflix TechBlog
Submissions from xiph.org Hacker News
Image formats for the web HEIC and AVIF – The Publishing Project
Comparing AVIF vs WebP file sizes at the same DSSIM
AV1 Codec
AVIF images color losschange AV1

Unfortunately I have not found a standard encoder with recommended settings. There appears to be zero guidance on settings anywhere. Because of that I am using the simplest encoder I could find, Kornelski's "cavif" which has a simple --quality parameter. I run thusly :

cavif --quality=X -o out.avif in.png
avifdef out.avif dec.png
imdiff in.png dec.png 
CALL FOR HELP : if you know better programs/settings for encoding AVIF, please let me know ; in avifenc , what is the min max Q parameter? adaptive quantization appears to be off by default, don't we want that on?

I measure results using my imdiff

I will compare AVIF to what I call "JPEG++" which is JPEG with the packjpg/Lepton back end, and a deblocking decoder (my "jpegdec"). This a rough stand-in for what I think a real JPEG++ should be (it should really have an R-D front-end and chroma-from-luma as well; that's all very easy and unequivocably good).

With no further ado, some results :

(imdiff "fit" is a quality in 0-10 , higher is better)

porsche640.bmp :

PDI_1200 :

Results are a bit disappointing. AVIF is much beter on RMSE but slightly worse on my other two scores. Overall that means it's most likely better overall, but it's not a huge margin.

(I'm sure AVIF is a big win on graphic/text images where JPEG does poorly)

AVIF results here look worse than what I saw from BPG (HEIC). Perhaps better encoders/settings will fix that.

Looking at the results visually, AVIF preserves sharp lines much better, but is completely throwing away detail in some places. There are some places where I see AVIF actually *change* the image, whereas JPEG is always just making the same image but slightly worse.

7z of encoded images to compare (2 MB)

NOTE : the fact that AVIF wins strongly in RGB RMSE but not in my other perceptual metrics indicates that is is not optimizing for those metrics. Perhaps in other perceptual metrics it would show a strong win. The metrics I use here from imdiff were chosen because I found them to be the best fit to human quality scores. Lots of the standard scores that people use (like naive SSIM) I have found to be utter junk, with no better correlation to human quality than RMSE. MS-SSIM-IW is the best variant of SSIM I know, but I haven't tested some of the newer metrics that have come out in the last few years.


Some JPEG Tools

A couple of tips and tools for working with JPEG files. I use :


Of course you can use exiftool and jpegtran but these are rather simpler.

1. Strip all personal data before uploading images to the web.

JPEG EXIF headers contain things like date shot and location. If you don't want Google to scrape that and use it to track your movements and send drones to follow you carrying advertisements, you might want to strip all that private data before you upload images to the net. I use :

jhead -purejpg %*
jhead -mkexif %*
jhead -dsft %*
with the very handy jhead tool. This removes all non-image headers, then makes a blank exif, then sets the exif time from the mod time. The last two steps are optional, if you want to preserve the shot time (assuming the mod time of the file was equal to the exif time). Note if the times are not equal you can use "jhead -ft" to set modtime from exif time before this.

Also note that if you use tools to modify images (resizing, changing exposure, what have you), they may or may not carry through the old exif data; whether that's good or bad is up to you, but it's very inconsistent. They also will probably set the modtime to now, whereas I prefer to keep the modtime = shot time so that I can find files by date more easily.

2. Remove orientation tags

iPhones are the only device I know of that consistently make use of the JPEG orientation tags rather than actually rotate the pixels. Unfortunately lots of loaders don't support these tags right, so if you load the image in a non-compliant loader it will appear sideways or upside down.

(this is a classic problem with data formats that have too many features; inevitabily many implementers only support a subset of the features which they deem to be the necessary ones; then some bone-head says "oh look the format has this weird feature, let's use it", and they output data which most loaders won't load right (then they blame the loaders). This is a mistake of the people defining the format; don't include unnecessary features, and make the practical subset a well-defined subset, not something that forms ad-hoc from use.)

To fix this, you want to read the orientation tag, rotate the pixels, then blank out the orientation tag (you have to do the last step or compliant loaders will doubly rotate). I used to have scripts to do all that with jpegtran, but it's super easy to do with jhead, you just :

jhead -autorot %*

3. Lossless crop

Avoid loading a JPEG, modifying it, and saving it out again. It destroys the image by re-applying the lossy quantization. I think by far the most common modification people do is cropping and there's no damn reason to re-encode for a crop. You can crop at 8x8 granularity losslessly, and you should. (all the web image uploaders that give you a crop tool need to do this, please, stop sucking so bad).

jpegcrop is a GUI tool that provides a decent lossless crop.

FreeVimager is okay too. Lossless crop is hidden in "JPEG Advanced".

We've got JPEG-XL coming soon, which looks great, but all the tech in the world won't help if people like Instagram & Youtube keep re-encoding uploads, and re-encoding every time you modify.

old rants