Skip to content

Instantly share code, notes, and snippets.

@Cyan4973

Cyan4973/stb_dxt.h

Forked from rygorous/stb_dxt.h
Created October 4, 2012 09:44
Show Gist options
  • Save Cyan4973/3832603 to your computer and use it in GitHub Desktop.
Save Cyan4973/3832603 to your computer and use it in GitHub Desktop.
Download ZIP
YannC's modified rygdxt code review
Raw
stb_dxt.h
// stb_dxt.h - Real-Time DXT1/DXT5 compressor
// Based on original by fabian "ryg" giesen v1.04
// Custom version, modified by Yann Collet
//
/*
BSD 2-Clause License (http://www.opensource.org/licenses/bsd-license.php)
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are
met:
* Redistributions of source code must retain the above copyright
notice, this list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above
copyright notice, this list of conditions and the following disclaimer
in the documentation and/or other materials provided with the
distribution.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
"AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
(INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
You can contact the author at :
- RygsDXTc source repository : http://code.google.com/p/rygsdxtc/
*/
// use '#define STB_DXT_IMPLEMENTATION' before including to create the implementation
//
// USAGE:
// call stb_compress_dxt_block() for every block (you must pad)
// source should be a 4x4 block of RGBA data in row-major order;
// A is ignored if you specify alpha=0; you can turn on dithering
// and "high quality" using mode.
//
// version history:
// v1.04 - (ryg) default to no rounding bias for lerped colors (as per S3TC/DX10 spec);
// single color match fix (allow for inexact color interpolation);
// optimal DXT5 index finder; "high quality" mode that runs multiple refinement steps.
// v1.03 - (stb) endianness support
// v1.02 - (stb) fix alpha encoding bug
// v1.01 - (stb) fix bug converting to RGB that messed up quality, thanks ryg & cbloom
// v1.00 - (stb) first release
#ifndef STB_INCLUDE_STB_DXT_H
#define STB_INCLUDE_STB_DXT_H
//*******************************************************************
// Enable custom Optimisations
// Comment this define if you want to revert to ryg's original code
#define NEW_OPTIMISATIONS
//*******************************************************************
// compression mode (bitflags)
#define STB_DXT_NORMAL 0
#define STB_DXT_DITHER 1 // use dithering. dubious win. never use for normal maps and the like!
#define STB_DXT_HIGHQUAL 2 // high quality mode, does two refinement steps instead of 1. ~30-40% slower.
void rygCompress( unsigned char *dst, unsigned char *src, int w, int h, int isDxt5 );
// TODO remove these, not working properly..
void rygCompressYCoCg( unsigned char *dst, unsigned char *src, int w, int h );
void linearize( unsigned char * dst, const unsigned char * src, int n );
void stb_compress_dxt_block(unsigned char *dest, const unsigned char *src, int alpha, int mode);
#define STB_COMPRESS_DXT_BLOCK
#ifdef STB_DXT_IMPLEMENTATION
// configuration options for DXT encoder. set them in the project/makefile or just define
// them at the top.
// STB_DXT_USE_ROUNDING_BIAS
// use a rounding bias during color interpolation. this is closer to what "ideal"
// interpolation would do but doesn't match the S3TC/DX10 spec. old versions (pre-1.03)
// implicitly had this turned on.
//
// in case you're targeting a specific type of hardware (e.g. console programmers):
// NVidia and Intel GPUs (as of 2010) as well as DX9 ref use DXT decoders that are closer
// to STB_DXT_USE_ROUNDING_BIAS. AMD/ATI, S3 and DX10 ref are closer to rounding with no bias.
// you also see "(a*5 + b*3) / 8" on some old GPU designs.
// #define STB_DXT_USE_ROUNDING_BIAS
#include <stdlib.h>
#include <math.h>
#include <stddef.h>
#include <string.h> // memset
#include <assert.h>
#include <iostream>
#include <algorithm>
static unsigned char stb__Expand5[32];
static unsigned char stb__Expand6[64];
static unsigned char stb__OMatch5[256][2];
static unsigned char stb__OMatch6[256][2];
static unsigned char stb__QuantRBTab[256+16];
static unsigned char stb__QuantGTab[256+16];
static int stb__Mul8Bit(int a, int b)
{
int t = a*b + 128;
return (t + (t >> 8)) >> 8;
}
static void stb__From16Bit(unsigned char *out, unsigned short v)
{
int rv = (v & 0xf800) >> 11;
int gv = (v & 0x07e0) >> 5;
int bv = (v & 0x001f) >> 0;
out[0] = stb__Expand5[rv];
out[1] = stb__Expand6[gv];
out[2] = stb__Expand5[bv];
out[3] = 0;
}
static unsigned short stb__As16Bit(int r, int g, int b)
{
return (stb__Mul8Bit(r,31) << 11) + (stb__Mul8Bit(g,63) << 5) + stb__Mul8Bit(b,31);
}
// linear interpolation at 1/3 point between a and b, using desired rounding type
static int stb__Lerp13(int a, int b)
{
#ifdef STB_DXT_USE_ROUNDING_BIAS
// with rounding bias
return a + stb__Mul8Bit(b-a, 0x55);
#else
// without rounding bias
// replace "/ 3" by "* 0xaaab) >> 17" if your compiler sucks or you really need every ounce of speed.
return (2*a + b) / 3;
#endif
}
// lerp RGB color
static void stb__Lerp13RGB(unsigned char *out, unsigned char *p1, unsigned char *p2)
{
out[0] = stb__Lerp13(p1[0], p2[0]);
out[1] = stb__Lerp13(p1[1], p2[1]);
out[2] = stb__Lerp13(p1[2], p2[2]);
}
/****************************************************************************/
// compute table to reproduce constant colors as accurately as possible
static void stb__PrepareOptTable(unsigned char *Table,const unsigned char *expand,int size)
{
int i,mn,mx;
for (i=0;i<256;i++) {
int bestErr = 256;
for (mn=0;mn<size;mn++) {
for (mx=0;mx<size;mx++) {
int mine = expand[mn];
int maxe = expand[mx];
int err = abs(stb__Lerp13(maxe, mine) - i);
// DX10 spec says that interpolation must be within 3% of "correct" result,
// add this as error term. (normally we'd expect a random distribution of
// +-1.5% error, but nowhere in the spec does it say that the error has to be
// unbiased - better safe than sorry).
err += abs(maxe - mine) * 3 / 100;
if(err < bestErr) {
Table[i*2+0] = mx;
Table[i*2+1] = mn;
bestErr = err;
}
}
}
}
}
static void stb__EvalColors(unsigned char *color,unsigned short c0,unsigned short c1)
{
stb__From16Bit(color+ 0, c0);
stb__From16Bit(color+ 4, c1);
stb__Lerp13RGB(color+ 8, color+0, color+4);
stb__Lerp13RGB(color+12, color+4, color+0);
}
// Block dithering function. Simply dithers a block to 565 RGB.
// (Floyd-Steinberg)
static void stb__DitherBlock(unsigned char *dest, unsigned char *block)
{
int err[8],*ep1 = err,*ep2 = err+4, *et;
int ch,y;
// process channels seperately
for (ch=0; ch<3; ++ch) {
unsigned char *bp = block+ch, *dp = dest+ch;
unsigned char *quant = (ch == 1) ? stb__QuantGTab+8 : stb__QuantRBTab+8;
memset(err, 0, sizeof(err));
for(y=0; y<4; ++y) {
dp[ 0] = quant[bp[ 0] + ((3*ep2[1] + 5*ep2[0]) >> 4)];
ep1[0] = bp[ 0] - dp[ 0];
dp[ 4] = quant[bp[ 4] + ((7*ep1[0] + 3*ep2[2] + 5*ep2[1] + ep2[0]) >> 4)];
ep1[1] = bp[ 4] - dp[ 4];
dp[ 8] = quant[bp[ 8] + ((7*ep1[1] + 3*ep2[3] + 5*ep2[2] + ep2[1]) >> 4)];
ep1[2] = bp[ 8] - dp[ 8];
dp[12] = quant[bp[12] + ((7*ep1[2] + 5*ep2[3] + ep2[2]) >> 4)];
ep1[3] = bp[12] - dp[12];
bp += 16;
dp += 16;
et = ep1, ep1 = ep2, ep2 = et; // swap
}
}
}
// The color matching function
static unsigned int stb__MatchColorsBlock(unsigned char *block, unsigned char *color,int dither)
{
unsigned int mask = 0;
int dirr = color[0*4+0] - color[1*4+0];
int dirg = color[0*4+1] - color[1*4+1];
int dirb = color[0*4+2] - color[1*4+2];
int dots[16];
int stops[4];
int i;
int c0Point, halfPoint, c3Point;
for(i=0;i<16;i++)
dots[i] = block[i*4+0]*dirr + block[i*4+1]*dirg + block[i*4+2]*dirb;
for(i=0;i<4;i++)
stops[i] = color[i*4+0]*dirr + color[i*4+1]*dirg + color[i*4+2]*dirb;
// think of the colors as arranged on a line; project point onto that line, then choose
// next color out of available ones. we compute the crossover points for "best color in top
// half"/"best in bottom half" and then the same inside that subinterval.
//
// relying on this 1d approximation isn't always optimal in terms of euclidean distance,
// but it's very close and a lot faster.
// http://cbloomrants.blogspot.com/2008/12/12-08-08-dxtc-summary.html
c0Point = (stops[1] + stops[3]) >> 1;
halfPoint = (stops[3] + stops[2]) >> 1;
c3Point = (stops[2] + stops[0]) >> 1;
if(!dither)
{
// the version without dithering is straightforward
#ifdef NEW_OPTIMISATIONS
// @@@ryg: Hmm, how much did this help?
// @@@cyan : about 7ms on a very large texture requiring 136ms for DXT1 conversion.
// i also tried another version with shift-to-left (<<) instead of (>>).
// the other version gained only 4ms.
const int indexMap[8] = { 0 << 30,2 << 30,0 << 30,2 << 30,3 << 30,3 << 30,1 << 30,1 << 30 };
for(int i=0;i<16;i++)
{
int dot = dots[i];
mask >>= 2;
int bits =( (dot < halfPoint) ? 4 : 0 )
| ( (dot < c0Point) ? 2 : 0 )
| ( (dot < c3Point) ? 1 : 0 );
mask |= indexMap[bits];
}
#else
for (i=15;i>=0;i--) {
int dot = dots[i];
mask <<= 2;
if(dot < halfPoint)
mask |= (dot < c0Point) ? 1 : 3;
else
mask |= (dot < c3Point) ? 2 : 0;
}
#endif
} else {
// with floyd-steinberg dithering
int err[8],*ep1 = err,*ep2 = err+4;
int *dp = dots, y;
c0Point <<= 4;
halfPoint <<= 4;
c3Point <<= 4;
for(i=0;i<8;i++)
err[i] = 0;
for(y=0;y<4;y++)
{
int dot,lmask,step;
dot = (dp[0] << 4) + (3*ep2[1] + 5*ep2[0]);
if(dot < halfPoint)
step = (dot < c0Point) ? 1 : 3;
else
step = (dot < c3Point) ? 2 : 0;
ep1[0] = dp[0] - stops[step];
lmask = step;
dot = (dp[1] << 4) + (7*ep1[0] + 3*ep2[2] + 5*ep2[1] + ep2[0]);
if(dot < halfPoint)
step = (dot < c0Point) ? 1 : 3;
else
step = (dot < c3Point) ? 2 : 0;
ep1[1] = dp[1] - stops[step];
lmask |= step<<2;
dot = (dp[2] << 4) + (7*ep1[1] + 3*ep2[3] + 5*ep2[2] + ep2[1]);
if(dot < halfPoint)
step = (dot < c0Point) ? 1 : 3;
else
step = (dot < c3Point) ? 2 : 0;
ep1[2] = dp[2] - stops[step];
lmask |= step<<4;
dot = (dp[3] << 4) + (7*ep1[2] + 5*ep2[3] + ep2[2]);
if(dot < halfPoint)
step = (dot < c0Point) ? 1 : 3;
else
step = (dot < c3Point) ? 2 : 0;
ep1[3] = dp[3] - stops[step];
lmask |= step<<6;
dp += 4;
mask |= lmask << (y*8);
{ int *et = ep1; ep1 = ep2; ep2 = et; } // swap
}
}
return mask;
}
// The color optimization function. (Clever code, part 1)
static void stb__OptimizeColorsBlock(unsigned char *block, unsigned short *pmax16, unsigned short *pmin16)
{
unsigned char *minp, *maxp;
double magn;
int v_r,v_g,v_b;
static const int nIterPower = 4;
float covf[6],vfr,vfg,vfb;
// determine color distribution
int cov[6];
int mu[3],min[3],max[3];
int ch,i,iter;
for(ch=0;ch<3;ch++)
{
const unsigned char *bp = ((const unsigned char *) block) + ch;
int muv,minv,maxv;
#ifdef NEW_OPTIMISATIONS
// @@@ryg: Really not sure about the unrolling here... normally, I'd prefer to "unroll" this
// by using SIMD instructions, if you need the speed.
// @@@cyan: the objective is not really unlooping,
// but rather favor ILP (Instruction Level Parallelism)
// by ensuring there is no dependancies between consecutive operations
// Furthermore, i don't know much about SIMD.
// I would like to keep the code in C, avoiding ASM as much as possible.
//
// @@@ryg:
// MIN/MAX pair:
// Your MIN is broken (it's actually MAX). Say a=1, b=2, and walk through MIN:
// (a-b) = -1, (a-b)>>31 = -1, (b-a) & -1 = 1 & -1 = 1, so MIN(a,b) = a + 1 = 2?! I don't think so!
// MIN should be: (sans casts)
// MIN(a,b) = a + ((a - b) & ((a - b) >> 31))
// this adds (a - b) to a if and only if (a - b) < 0, which is what you want.
//
// @@@cyan: you are correct, the macro was broken.
// reversing the mask condition was enough to correct the problem. Thanks for suggestion.
// by the way, it's curious that such a blatant mistake did not result in any visible artefact.
// if i do understand properly, the older version was swapping MIN & MAX.
//
// @@@ryg:
// So assume you fix this. Then the MAX form you use still performs extra work if you want to
// compute both MIN and MAX simultaneously. You want to share computations where possible, so
// a better form in this context is:
// MAX(a,b) = b - ((a - b) & ((a - b) >> 31))
//
// Then for something like MINMAX, you would get something like the following code: (x86)
//
// ; in: eax = a, ebx = b, ecx/edx = scratch
// mov ecx, eax
// sub ecx, ebx
// mov edx, ecx
// sar ecx, 31
// and ecx, edx ; (a-b) & ((a-b) >> 31)
// add eax, ecx ; now eax=min(a,b)
// sub ebx, ecx ; now ebx=max(a,b)
//
// i.e. the MAX is basically free in that pair. This is related to the (a+b)-MIN(a,b) trick
// you use in RANGE.
//
// @@@cyan: not really.
// notice that it's not about the min and max of a & b
// it's about the min of min##a and min##b, and the max of max##a et max##b
// so they are 2 different set of values
//
// The version below is the old one, for reference.
// The new, corrected, version will be updated at google code.
//
# define MIN(a,b) (int)a + ( ((int)b-a) & ( ((int)a-b) >> 31 ) )
# define MAX(a,b) (int)a + ( ((int)b-a) & ( ((int)b-a) >> 31 ) )
# define RANGE(a,b,n) int total##n = a+b; int min##n = MIN(a,b); int max##n = total##n - min##n;
# define MINMAX(a,b,n) int min##n = MIN(min##a, min##b); int max##n = MAX(max##a, max##b);
RANGE(bp[0], bp[4], 1);
muv = total1;
RANGE(bp[8], bp[12], 2);
muv += total2;
RANGE(bp[16], bp[20], 3);
muv += total3;
RANGE(bp[24], bp[28], 4);
muv += total4;
RANGE(bp[32], bp[36], 5);
muv += total5;
RANGE(bp[40], bp[44], 6);
muv += total6;
RANGE(bp[48], bp[52], 7);
muv += total7;
RANGE(bp[56], bp[60], 8);
muv += total8;
MINMAX(1,2,9);
MINMAX(3,4,10);
MINMAX(5,6,11);
MINMAX(7,8,12);
MINMAX(9,10,13);
MINMAX(11,12,14);
minv = MIN(min13,min14);
maxv = MAX(max13,max14);
#else
muv = minv = maxv = bp[0];
for(i=4;i<64;i+=4)
{
muv += bp[i];
if (bp[i] < minv) minv = bp[i];
else if (bp[i] > maxv) maxv = bp[i];
}
#endif
mu[ch] = (muv + 8) >> 4;
min[ch] = minv;
max[ch] = maxv;
}
// determine covariance matrix
for (i=0;i<6;i++)
cov[i] = 0;
for (i=0;i<16;i++)
{
int r = block[i*4+0] - mu[0];
int g = block[i*4+1] - mu[1];
int b = block[i*4+2] - mu[2];
cov[0] += r*r;
cov[1] += r*g;
cov[2] += r*b;
cov[3] += g*g;
cov[4] += g*b;
cov[5] += b*b;
}
// convert covariance matrix to float, find principal axis via power iter
for(i=0;i<6;i++)
covf[i] = cov[i] / 255.0f;
vfr = (float) (max[0] - min[0]);
vfg = (float) (max[1] - min[1]);
vfb = (float) (max[2] - min[2]);
for(iter=0;iter<nIterPower;iter++)
{
float r = vfr*covf[0] + vfg*covf[1] + vfb*covf[2];
float g = vfr*covf[1] + vfg*covf[3] + vfb*covf[4];
float b = vfr*covf[2] + vfg*covf[4] + vfb*covf[5];
vfr = r;
vfg = g;
vfb = b;
}
magn = fabs(vfr);
if (fabs(vfg) > magn) magn = fabs(vfg);
if (fabs(vfb) > magn) magn = fabs(vfb);
if(magn < 4.0f)
{ // too small, default to luminance
v_r = 299; // JPEG YCbCr luma coefs, scaled by 1000.
v_g = 587;
v_b = 114;
} else {
magn = 512.0 / magn;
v_r = (int) (vfr * magn);
v_g = (int) (vfg * magn);
v_b = (int) (vfb * magn);
}
#ifdef NEW_OPTIMISATIONS
// @@@ryg Did you verify that getting rid of this one compare was worthwhile?
// @@@cyan : i'm supposed to have checked each modification "worthiness" (speedwise)
// but strangely, i don't remember this one. I'll check again.
//
// Pick colors at extreme points
int mind, maxd;
mind = maxd = block[0]*v_r + block[1]*v_g + block[2]*v_b;
minp = maxp = block;
for(i=1;i<16;i++)
{
int dot = block[i*4+0]*v_r + block[i*4+1]*v_g + block[i*4+2]*v_b;
if (dot < mind) {
mind = dot;
minp = block+i*4;
continue;
}
if (dot > maxd) {
maxd = dot;
maxp = block+i*4;
}
}
#else
int mind = 0x7fffffff,maxd = -0x7fffffff;
// Pick colors at extreme points
for(i=0;i<16;i++)
{
int dot = block[i*4+0]*v_r + block[i*4+1]*v_g + block[i*4+2]*v_b;
if (dot < mind) {
mind = dot;
minp = block+i*4;
}
if (dot > maxd) {
maxd = dot;
maxp = block+i*4;
}
}
#endif
*pmax16 = stb__As16Bit(maxp[0],maxp[1],maxp[2]);
*pmin16 = stb__As16Bit(minp[0],minp[1],minp[2]);
}
inline static int stb__sclamp(float y, int p0, int p1)
{
int x = (int) y;
#ifdef NEW_OPTIMISATIONS
x = x>p1 ? p1 : x;
return x<p0 ? p0 : x;
#else
if (x < p0) return p0;
if (x > p1) return p1;
return x;
#endif
}
// The refinement function. (Clever code, part 2)
// Tries to optimize colors to suit block contents better.
// (By solving a least squares system via normal equations+Cramer's rule)
static int stb__RefineBlock(unsigned char *block, unsigned short *pmax16, unsigned short *pmin16, unsigned int mask)
{
static const int w1Tab[4] = { 3,0,2,1 };
static const int prods[4] = { 0x090000,0x000900,0x040102,0x010402 };
// ^some magic to save a lot of multiplies in the accumulating loop...
// (precomputed products of weights for least squares system, accumulated inside one 32-bit register)
float frb,fg;
unsigned short oldMin, oldMax, min16, max16;
int i, akku = 0, xx,xy,yy;
int At1_r,At1_g,At1_b;
int At2_r,At2_g,At2_b;
unsigned int cm = mask;
oldMin = *pmin16;
oldMax = *pmax16;
if((mask ^ (mask<<2)) < 4) // all pixels have the same index?
{
// yes, linear system would be singular; solve using optimal
// single-color match on average color
int r = 8, g = 8, b = 8;
for (i=0;i<16;++i) {
r += block[i*4+0];
g += block[i*4+1];
b += block[i*4+2];
}
r >>= 4; g >>= 4; b >>= 4;
max16 = (stb__OMatch5[r][0]<<11) | (stb__OMatch6[g][0]<<5) | stb__OMatch5[b][0];
min16 = (stb__OMatch5[r][1]<<11) | (stb__OMatch6[g][1]<<5) | stb__OMatch5[b][1];
} else {
At1_r = At1_g = At1_b = 0;
At2_r = At2_g = At2_b = 0;
for (i=0;i<16;++i,cm>>=2)
{
int step = cm&3;
int w1 = w1Tab[step];
int r = block[i*4+0];
int g = block[i*4+1];
int b = block[i*4+2];
akku += prods[step];
At1_r += w1*r;
At1_g += w1*g;
At1_b += w1*b;
At2_r += r;
At2_g += g;
At2_b += b;
}
At2_r = 3*At2_r - At1_r;
At2_g = 3*At2_g - At1_g;
At2_b = 3*At2_b - At1_b;
// extract solutions and decide solvability
xx = akku >> 16;
yy = (akku >> 8) & 0xff;
xy = (akku >> 0) & 0xff;
frb = 3.0f * 31.0f / 255.0f / (xx*yy - xy*xy);
fg = frb * 63.0f / 31.0f;
// solve.
max16 = stb__sclamp((At1_r*yy - At2_r*xy)*frb+0.5f,0,31) << 11;
max16 |= stb__sclamp((At1_g*yy - At2_g*xy)*fg +0.5f,0,63) << 5;
max16 |= stb__sclamp((At1_b*yy - At2_b*xy)*frb+0.5f,0,31) << 0;
min16 = stb__sclamp((At2_r*xx - At1_r*xy)*frb+0.5f,0,31) << 11;
min16 |= stb__sclamp((At2_g*xx - At1_g*xy)*fg +0.5f,0,63) << 5;
min16 |= stb__sclamp((At2_b*xx - At1_b*xy)*frb+0.5f,0,31) << 0;
}
*pmin16 = min16;
*pmax16 = max16;
return oldMin != min16 || oldMax != max16;
}
// Color block compression
static void stb__CompressColorBlock(unsigned char *dest, unsigned char *block, int mode)
{
unsigned int mask;
int i;
int dither;
int refinecount;
unsigned short max16, min16;
unsigned char dblock[16*4],color[4*4];
dither = mode & STB_DXT_DITHER;
refinecount = (mode & STB_DXT_HIGHQUAL) ? 2 : 1;
// check if block is constant
for (i=1;i<16;i++)
if (((unsigned int *) block)[i] != ((unsigned int *) block)[0])
break;
if(i == 16)
{ // constant color
int r = block[0], g = block[1], b = block[2];
mask = 0xaaaaaaaa;
max16 = (stb__OMatch5[r][0]<<11) | (stb__OMatch6[g][0]<<5) | stb__OMatch5[b][0];
min16 = (stb__OMatch5[r][1]<<11) | (stb__OMatch6[g][1]<<5) | stb__OMatch5[b][1];
} else
{
// first step: compute dithered version for PCA if desired
if(dither)
stb__DitherBlock(dblock,block);
// second step: pca+map along principal axis
stb__OptimizeColorsBlock(dither ? dblock : block,&max16,&min16);
if (max16 != min16)
{
stb__EvalColors(color,max16,min16);
mask = stb__MatchColorsBlock(block,color,dither);
} else
mask = 0;
// third step: refine (multiple times if requested)
for (i=0;i<refinecount;i++) {
unsigned int lastmask = mask;
if (stb__RefineBlock(dither ? dblock : block,&max16,&min16,mask))
{
if (max16 != min16)
{
stb__EvalColors(color,max16,min16);
mask = stb__MatchColorsBlock(block,color,dither);
} else
{
mask = 0;
break;
}
}
if(mask == lastmask)
break;
}
}
// write the color block
if(max16 < min16)
{
unsigned short t = min16;
min16 = max16;
max16 = t;
mask ^= 0x55555555;
}
dest[0] = (unsigned char) (max16);
dest[1] = (unsigned char) (max16 >> 8);
dest[2] = (unsigned char) (min16);
dest[3] = (unsigned char) (min16 >> 8);
dest[4] = (unsigned char) (mask);
dest[5] = (unsigned char) (mask >> 8);
dest[6] = (unsigned char) (mask >> 16);
dest[7] = (unsigned char) (mask >> 24);
}
// Alpha block compression (this is easy for a change)
static void stb__CompressAlphaBlock(unsigned char *dest,unsigned char *src,int mode)
{
int i,dist,bias,dist4,dist2,bits,mask;
// find min/max color
int mn,mx;
mn = mx = src[3];
for (i=1;i<16;i++)
{
if (src[i*4+3] < mn) mn = src[i*4+3];
else if (src[i*4+3] > mx) mx = src[i*4+3];
}
// encode them
((unsigned char *)dest)[0] = mx;
((unsigned char *)dest)[1] = mn;
dest += 2;
#ifdef NEW_OPTIMISATIONS
// @@@ryg I like the shortcut idea, but this test should be "mn==mx" instead.
// (So it will catch *any* block with only one alpha value)
// @@cyan : the index pattern following alpha values is applicable for opaque alpha.
// but i'm not sure if it is applicable to any uniform alpha (?)
// opaque shortcut
if (mn==255)
{
((unsigned char *)dest)[2] = 0x49;
((unsigned char *)dest)[3] = 0x92;
((unsigned char *)dest)[4] = 0x24;
((unsigned char *)dest)[5] = 0x49;
((unsigned char *)dest)[6] = 0x92;
((unsigned char *)dest)[7] = 0x24;
return;
}
#endif
// determine bias and emit color indices
// given the choice of mx/mn, these indices are optimal:
// http://fgiesen.wordpress.com/2009/12/15/dxt5-alpha-block-index-determination/
dist = mx-mn;
//printf("mn = %i; mx = %i; dist = %i\n", mn, mx, dist);
dist4 = dist*4;
dist2 = dist*2;
bias = (dist < 8) ? (dist - 1) : (dist/2 + 2);
bias -= mn * 7;
bits = 0, mask=0;
for (i=0;i<16;i++)
{
int a = src[i*4+3]*7 + bias;
int ind,t;
// select index. this is a "linear scale" lerp factor between 0 (val=min) and 7 (val=max).
t = (a >= dist4) ? -1 : 0; ind = t & 4; a -= dist4 & t;
t = (a >= dist2) ? -1 : 0; ind += t & 2; a -= dist2 & t;
ind += (a >= dist);
// turn linear scale into DXT index (0/1 are extremal pts)
ind = -ind & 7;
ind ^= (2 > ind);
// write index
mask |= ind << bits;
if((bits += 3) >= 8)
{
*dest++ = mask;
mask >>= 8;
bits -= 8;
}
}
}
static void stb__InitDXT()
{
int i;
for(i=0;i<32;i++)
stb__Expand5[i] = (i<<3)|(i>>2);
for(i=0;i<64;i++)
stb__Expand6[i] = (i<<2)|(i>>4);
for(i=0;i<256+16;i++)
{
int v = i-8 < 0 ? 0 : i-8 > 255 ? 255 : i-8;
stb__QuantRBTab[i] = stb__Expand5[stb__Mul8Bit(v,31)];
stb__QuantGTab[i] = stb__Expand6[stb__Mul8Bit(v,63)];
}
stb__PrepareOptTable(&stb__OMatch5[0][0],stb__Expand5,32);
stb__PrepareOptTable(&stb__OMatch6[0][0],stb__Expand6,64);
}
void stb_compress_dxt_block(unsigned char *dest, const unsigned char *src, int alpha, int mode)
{
static int init=1;
if (init)
{
stb__InitDXT();
init=0;
}
if (alpha)
{
stb__CompressAlphaBlock(dest,(unsigned char*) src,mode);
dest += 8;
}
stb__CompressColorBlock(dest,(unsigned char*) src,mode);
}
int imin(int x, int y) { return (x < y) ? x : y; }
static void extractBlock(const unsigned char *src, int x, int y,
int w, int h, unsigned char *block)
{
int i, j;
#ifdef NEW_OPTIMISATIONS
// @@@ryg: Yep, you want to special-case full-size blocks.
if ((w-x >=4) && (h-y >=4))
{
// Full Square shortcut
src += x*4;
src += y*w*4;
for (i=0; i < 4; ++i)
{
*(unsigned int*)block = *(unsigned int*) src; block += 4; src += 4;
*(unsigned int*)block = *(unsigned int*) src; block += 4; src += 4;
*(unsigned int*)block = *(unsigned int*) src; block += 4; src += 4;
*(unsigned int*)block = *(unsigned int*) src; block += 4;
src += (w*4) - 12;
}
return;
}
#endif
int bw = imin(w - x, 4);
int bh = imin(h - y, 4);
int bx, by;
const int rem[] =
{
0, 0, 0, 0,
0, 1, 0, 1,
0, 1, 2, 0,
0, 1, 2, 3
};
for(i = 0; i < 4; ++i)
{
by = rem[(bh - 1) * 4 + i] + y;
for(j = 0; j < 4; ++j)
{
bx = rem[(bw - 1) * 4 + j] + x;
block[(i * 4 * 4) + (j * 4) + 0] =
src[(by * (w * 4)) + (bx * 4) + 0];
block[(i * 4 * 4) + (j * 4) + 1] =
src[(by * (w * 4)) + (bx * 4) + 1];
block[(i * 4 * 4) + (j * 4) + 2] =
src[(by * (w * 4)) + (bx * 4) + 2];
block[(i * 4 * 4) + (j * 4) + 3] =
src[(by * (w * 4)) + (bx * 4) + 3];
}
}
}
// should be a pretty optimized 0-255 clamper
inline static unsigned char clamp255( int n )
{
if( n > 255 ) n = 255;
if( n < 0 ) n = 0;
return n;
}
void rgbToYCoCgBlock( unsigned char * dst, const unsigned char * src )
{
// Calculate Co and Cg extents
int extents = 0;
int n = 0;
int iY, iCo, iCg; //, r, g, b;
int blockCo[16];
int blockCg[16];
int i;
const unsigned char *px = src;
for(i=0;i<n;i++)
{
iCo = (px[0]<<1) - (px[2]<<1);
iCg = (px[1]<<1) - px[0] - px[2];
if(-iCo > extents) extents = -iCo;
if( iCo > extents) extents = iCo;
if(-iCg > extents) extents = -iCg;
if( iCg > extents) extents = iCg;
blockCo[n] = iCo;
blockCg[n++] = iCg;
px += 4;
}
// Co = -510..510
// Cg = -510..510
float scaleFactor = 1.0f;
if(extents > 127)
scaleFactor = (float)extents * 4.0f / 510.0f;
// Convert to quantized scalefactor
unsigned char scaleFactorQuantized = (unsigned char)(ceil((scaleFactor - 1.0f) * 31.0f / 3.0f));
// Unquantize
scaleFactor = 1.0f + (float)(scaleFactorQuantized / 31.0f) * 3.0f;
unsigned char bVal = (unsigned char)((scaleFactorQuantized << 3) | (scaleFactorQuantized >> 2));
unsigned char *outPx = dst;
n = 0;
px = src;
/*
for(i=0;i<16;i++)
{
// Calculate components
iY = ( px[0] + (px[1]<<1) + px[2] + 2 ) / 4;
iCo = ((blockCo[n] / scaleFactor) + 128);
iCg = ((blockCg[n] / scaleFactor) + 128);
if(iCo < 0) iCo = 0; else if(iCo > 255) iCo = 255;
if(iCg < 0) iCg = 0; else if(iCg > 255) iCg = 255;
if(iY < 0) iY = 0; else if(iY > 255) iY = 255;
px += 4;
outPx[0] = (unsigned char)iCo;
outPx[1] = (unsigned char)iCg;
outPx[2] = bVal;
outPx[3] = (unsigned char)iY;
outPx += 4;
}*/
for(i=0;i<16;i++)
{
// Calculate components
int r = px[0];
int g = (px[1] + 1) >> 1;
int b = px[2];
int tmp = (2 + r + b) >> 2;
// Co
iCo = clamp255( 128 + ((r - b + 1) >> 1) );
// Y
iY = clamp255( g + tmp );
// Cg
iCg = clamp255( 128 + g - tmp );
px += 4;
outPx[0] = (unsigned char)iCo;
outPx[1] = (unsigned char)iCg;
outPx[2] = bVal;
outPx[3] = (unsigned char)iY;
outPx += 4;
}
}
void rygCompress( unsigned char *dst, unsigned char *src, int w, int h, int isDxt5 )
{
unsigned char block[64];
int x, y;
for(y = 0; y < h; y += 4)
{
for(x = 0; x < w; x += 4)
{
extractBlock(src, x, y, w, h, block);
stb_compress_dxt_block(dst, block, isDxt5, 10);
dst += isDxt5 ? 16 : 8;
}
}
}
void rygCompressYCoCg( unsigned char *dst, unsigned char *src, int w, int h )
{
unsigned char block[64];
unsigned char ycocgblock[64];
int x, y;
for(y = 0; y < h; y += 4)
{
for(x = 0; x < w; x += 4)
{
extractBlock(src, x, y, w, h, block);
rgbToYCoCgBlock(ycocgblock,block);
stb_compress_dxt_block(dst, ycocgblock, 1, 10);
dst += 16;
}
}
}
static void stbgl__compress(unsigned char *p, unsigned char *rgba, int w, int h, int isDxt5)
{
int i,j,y,y2;
int alpha = isDxt5;
for (j=0; j < w; j += 4) {
int x=4;
for (i=0; i < h; i += 4) {
unsigned char block[16*4];
if (i+3 >= w) x = w-i;
for (y=0; y < 4; ++y) {
if (j+y >= h) break;
memcpy(block+y*16, rgba + w*4*(j+y) + i*4, x*4);
}
if (x < 4) {
switch (x) {
case 0: assert(0);
case 1:
for (y2=0; y2 < y; ++y2) {
memcpy(block+y2*16+1*4, block+y2*16+0*4, 4);
memcpy(block+y2*16+2*4, block+y2*16+0*4, 8);
}
break;
case 2:
for (y2=0; y2 < y; ++y2)
memcpy(block+y2*16+2*4, block+y2*16+0*4, 8);
break;
case 3:
for (y2=0; y2 < y; ++y2)
memcpy(block+y2*16+3*4, block+y2*16+1*4, 4);
break;
}
}
y2 = 0;
for(; y<4; ++y,++y2)
memcpy(block+y*16, block+y2*16, 4*4);
stb_compress_dxt_block(p, block, alpha, 10);
p += alpha ? 16 : 8;
}
}
// assert(p <= end);
}
static inline unsigned char linearize(unsigned char inByte)
{
float srgbVal = ((float)inByte) / 255.0f;
float linearVal;
if(srgbVal < 0.04045)
linearVal = srgbVal / 12.92f;
else
linearVal = pow( (srgbVal + 0.055f) / 1.055f, 2.4f);
return (unsigned char)(floor(sqrt(linearVal)* 255.0 + 0.5));
}
void linearize( unsigned char * dst, const unsigned char * src, int n )
{
n*=4;
for( int i = 0; i < n; i++ )
dst[i] = linearize(src[i]);
}
#endif // STB_DXT_IMPLEMENTATION
#endif // STB_INCLUDE_STB_DXT_H
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment