drewcassidy/nvidia-texture-tools
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

osx fixes. Fix issue 211.

Browse Source
This commit is contained in:
castano
2014-12-02 20:23:21 +00:00
parent 2d6fc0e304
commit 7e2a9d1adb
42 changed files with 176 additions and 940 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
src/bc6h/CMakeLists.txt Normal file
View File

@ -0,0 +1,22 @@
PROJECT(bc6h)
INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
SET(BC6H_SRCS
bits.h
shapes_two.h
tile.h
zoh_utils.cpp
zoh_utils.h
zoh.cpp
zoh.h
zohone.cpp
zohtwo.cpp)
ADD_LIBRARY(bc6h STATIC ${BC6H_SRCS})
IF(NOT WIN32)
IF(CMAKE_COMPILER_IS_GNUCXX)
SET_TARGET_PROPERTIES(bc6h PROPERTIES COMPILE_FLAGS -fPIC)
ENDIF(CMAKE_COMPILER_IS_GNUCXX)
ENDIF(NOT WIN32)

76
src/bc6h/bits.h Normal file
View File

@ -0,0 +1,76 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
#pragma once
#ifndef _ZOH_BITS_H
#define _ZOH_BITS_H
// read/write a bitstream
#include "nvcore/Debug.h"
namespace ZOH {
class Bits
{
public:
Bits(char *data, int maxdatabits) { nvAssert (data && maxdatabits > 0); bptr = bend = 0; bits = data; maxbits = maxdatabits; readonly = 0;}
Bits(const char *data, int availdatabits) { nvAssert (data && availdatabits > 0); bptr = 0; bend = availdatabits; cbits = data; maxbits = availdatabits; readonly = 1;}
void write(int value, int nbits) {
nvAssert (nbits >= 0 && nbits < 32);
nvAssert (sizeof(int)>= 4);
for (int i=0; i<nbits; ++i)
writeone(value>>i);
}
int read(int nbits) {
nvAssert (nbits >= 0 && nbits < 32);
nvAssert (sizeof(int)>= 4);
int out = 0;
for (int i=0; i<nbits; ++i)
out |= readone() << i;
return out;
}
int getptr() { return bptr; }
void setptr(int ptr) { nvAssert (ptr >= 0 && ptr < maxbits); bptr = ptr; }
int getsize() { return bend; }
private:
int bptr; // next bit to read
int bend; // last written bit + 1
char *bits; // ptr to user bit stream
const char *cbits; // ptr to const user bit stream
int maxbits; // max size of user bit stream
char readonly; // 1 if this is a read-only stream
int readone() {
nvAssert (bptr < bend);
if (bptr >= bend) return 0;
int bit = (readonly ? cbits[bptr>>3] : bits[bptr>>3]) & (1 << (bptr & 7));
++bptr;
return bit != 0;
}
void writeone(int bit) {
nvAssert (!readonly); // "Writing a read-only bit stream"
nvAssert (bptr < maxbits);
if (bptr >= maxbits) return;
if (bit&1)
bits[bptr>>3] |= 1 << (bptr & 7);
else
bits[bptr>>3] &= ~(1 << (bptr & 7));
if (bptr++ >= bend) bend = bptr;
}
};
}
#endif

133
src/bc6h/shapes_two.h Normal file
View File

@ -0,0 +1,133 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
#pragma once
#ifndef _ZOH_SHAPES_TWO_H
#define _ZOH_SHAPES_TWO_H
// shapes for two regions
#define NREGIONS 2
#define NSHAPES 64
#define SHAPEBITS 6
static const int shapes[NSHAPES*16] =
{
0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1, 0, 0, 0, 1,
0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1,
0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1,
0, 0, 1, 1, 0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 1,
0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 1, 0, 1, 1, 1, 0, 0, 1, 1, 0, 0, 0, 1,
0, 0, 0, 1, 0, 1, 1, 1, 0, 1, 1, 1, 0, 0, 1, 1,
0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1,
0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 0,
0, 0, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 0, 0, 1,
0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1,
0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 0, 0, 0, 0,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 1, 1, 1,
1, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1, 1,
1, 1, 1, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 1,
1, 1, 1, 1, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0,
0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1,
0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1,
0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1,
0, 0, 0, 0, 1, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0, 1,
0, 0, 1, 1, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1,
0, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 1, 1, 0,
0, 0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 0, 0, 1, 1, 0,
0, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0,
0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 1,
0, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 0, 0, 1,
1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 1,
1, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 1, 1, 0, 0,
0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0, 1, 1,
0, 1, 0, 1, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0, 1, 1,
0, 1, 0, 1, 0, 0, 0, 0, 0, 1, 0, 1, 1, 1, 0, 0,
0, 1, 0, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 0, 0,
0, 0, 1, 1, 0, 1, 0, 1, 0, 1, 1, 0, 0, 1, 0, 1,
1, 1, 0, 0, 0, 1, 0, 1, 1, 0, 0, 1, 1, 0, 1, 0,
0, 0, 1, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 0, 1, 0,
1, 1, 0, 0, 1, 0, 1, 0, 1, 0, 0, 1, 0, 1, 0, 1,
0, 1, 1, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0, 0, 1, 1,
0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 0, 1, 0, 1, 1,
1, 1, 0, 0, 1, 1, 0, 0, 0, 1, 0, 0, 1, 1, 0, 1,
1, 1, 1, 0, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 0, 0,
0, 1, 1, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 0, 0,
1, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 0, 0, 1, 1, 0,
1, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, 0, 1, 1, 0,
0, 1, 1, 0, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0,
0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, 0, 0,
0, 1, 0, 0, 0, 0, 1, 0, 0, 1, 1, 1, 1, 1, 1, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 1, 0, 0,
0, 1, 1, 0, 0, 0, 1, 1, 0, 1, 1, 0, 0, 0, 1, 1,
1, 1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 1, 1, 0, 0, 1,
1, 0, 0, 1, 1, 1, 0, 0, 1, 0, 0, 1, 1, 1, 0, 0,
0, 0, 1, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 0,
0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1, 1, 0, 0, 0, 1,
1, 1, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 1, 0, 0, 0,
1, 1, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 1, 1, 1, 0,
1, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 1, 0, 1, 1, 1,
0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 0, 0,
1, 1, 1, 1, 0, 0, 1, 1, 0, 0, 1, 0, 0, 1, 0, 0,
0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 1,
0, 0, 1, 1, 0, 0, 0, 0, 1, 1, 1, 0, 0, 1, 1, 1,
};
#define REGION(x,y,si) shapes[((si)&3)*4+((si)>>2)*64+(x)+(y)*16]
static const int shapeindex_to_compressed_indices[NSHAPES*2] =
{
0,15, 0,15, 0,15, 0,15,
0,15, 0,15, 0,15, 0,15,
0,15, 0,15, 0,15, 0,15,
0,15, 0,15, 0,15, 0,15,
0,15, 0, 2, 0, 8, 0, 2,
0, 2, 0, 8, 0, 8, 0,15,
0, 2, 0, 8, 0, 2, 0, 2,
0, 8, 0, 8, 0, 2, 0, 2,
0,15, 0,15, 0, 6, 0, 8,
0, 2, 0, 8, 0,15, 0,15,
0, 2, 0, 8, 0, 2, 0, 2,
0, 2, 0,15, 0,15, 0, 6,
0, 6, 0, 2, 0, 6, 0, 8,
0,15, 0,15, 0, 2, 0, 2,
0,15, 0,15, 0,15, 0,15,
0,15, 0, 2, 0, 2, 0,15
};
#define SHAPEINDEX_TO_COMPRESSED_INDICES(si,region) shapeindex_to_compressed_indices[(si)*2+(region)]
#endif

83
src/bc6h/tile.h Normal file
View File

@ -0,0 +1,83 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
#pragma once
#ifndef _ZOH_TILE_H
#define _ZOH_TILE_H
#include "zoh_utils.h"
#include "nvmath/Vector.h"
#include <math.h>
namespace ZOH {
//#define USE_IMPORTANCE_MAP 1 // define this if you want to increase importance of some pixels in tile
class Tile
{
public:
// NOTE: this returns the appropriately-clamped BIT PATTERN of the half as an INTEGRAL float value
static float half2float(uint16 h)
{
return (float) Utils::ushort_to_format(h);
}
// NOTE: this is the inverse of the above operation
static uint16 float2half(float f)
{
return Utils::format_to_ushort((int)f);
}
// look for adjacent pixels that are identical. if there are enough of them, increase their importance
void generate_importance_map()
{
// initialize
for (int y=0; y<size_y; ++y)
for (int x=0; x<size_x; ++x)
{
// my importance is increased if I am identical to any of my 4-neighbors
importance_map[y][x] = match_4_neighbor(x,y) ? 5.0f : 1.0f;
}
}
bool is_equal(int x, int y, int xn, int yn)
{
if (xn < 0 || xn >= size_x || yn < 0 || yn >= size_y)
return false;
return( (data[y][x].x == data[yn][xn].x) &&
(data[y][x].y == data[yn][xn].y) &&
(data[y][x].z == data[yn][xn].z) );
}
#ifdef USE_IMPORTANCE_MAP
bool match_4_neighbor(int x, int y)
{
return is_equal(x,y,x-1,y) || is_equal(x,y,x+1,y) || is_equal(x,y,x,y-1) || is_equal(x,y,x,y+1);
}
#else
bool match_4_neighbor(int x, int y)
{
return false;
}
#endif
Tile() {};
~Tile(){};
Tile(int xs, int ys) {size_x = xs; size_y = ys;}
static const int TILE_H = 4;
static const int TILE_W = 4;
static const int TILE_TOTAL = TILE_H * TILE_W;
nv::Vector3 data[TILE_H][TILE_W];
float importance_map[TILE_H][TILE_W];
int size_x, size_y; // actual size of tile
};
}
#endif // _ZOH_TILE_H

197
src/bc6h/zoh.cpp Normal file
View File

@ -0,0 +1,197 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
// the zoh compressor and decompressor
#include "tile.h"
#include "zoh.h"
#include <string.h> // memcpy
using namespace ZOH;
bool ZOH::isone(const char *block)
{
char code = block[0] & 0x1F;
return (code == 0x03 || code == 0x07 || code == 0x0b || code == 0x0f);
}
void ZOH::compress(const Tile &t, char *block)
{
char oneblock[ZOH::BLOCKSIZE], twoblock[ZOH::BLOCKSIZE];
float mseone = ZOH::compressone(t, oneblock);
float msetwo = ZOH::compresstwo(t, twoblock);
if (mseone <= msetwo)
memcpy(block, oneblock, ZOH::BLOCKSIZE);
else
memcpy(block, twoblock, ZOH::BLOCKSIZE);
}
void ZOH::decompress(const char *block, Tile &t)
{
if (ZOH::isone(block))
ZOH::decompressone(block, t);
else
ZOH::decompresstwo(block, t);
}
/*
void ZOH::compress(string inf, string zohf)
{
Array2D<Rgba> pixels;
int w, h;
char block[ZOH::BLOCKSIZE];
Exr::readRgba(inf, pixels, w, h);
FILE *zohfile = fopen(zohf.c_str(), "wb");
if (zohfile == NULL) throw "Unable to open .zoh file for write";
// stuff for progress bar O.o
int ntiles = ((h+Tile::TILE_H-1)/Tile::TILE_H)*((w+Tile::TILE_W-1)/Tile::TILE_W);
int tilecnt = 0;
int ndots = 25;
int dotcnt = 0;
printf("Progress [");
for (int i=0; i<ndots;++i) printf(" ");
printf("]\rProgress ["); fflush(stdout);
// convert to tiles and compress each tile
for (int y=0; y<h; y+=Tile::TILE_H)
{
int ysize = min(Tile::TILE_H, h-y);
for (int x=0; x<w; x+=Tile::TILE_W)
{
int xsize = min(Tile::TILE_W, w-x);
Tile t(xsize, ysize);
t.insert(pixels, x, y);
ZOH::compress(t, block);
if (fwrite(block, sizeof(char), ZOH::BLOCKSIZE, zohfile) != ZOH::BLOCKSIZE)
throw "File error on write";
// progress bar
++tilecnt;
if (tilecnt > (ntiles * dotcnt)/ndots) { printf("."); fflush(stdout); ++dotcnt; }
}
}
printf("]\n"); // advance to next line finally
if (fclose(zohfile)) throw "Close failed on .zoh file";
}
static int str2int(std::string s)
{
int thing;
std::stringstream str (stringstream::in | stringstream::out);
str << s;
str >> thing;
return thing;
}
// zoh file name is ...-w-h.zoh, extract width and height
static void extract(string zohf, int &w, int &h)
{
size_t n = zohf.rfind('.', zohf.length()-1);
size_t n1 = zohf.rfind('-', n-1);
size_t n2 = zohf.rfind('-', n1-1);
string width = zohf.substr(n2+1, n1-n2-1);
w = str2int(width);
string height = zohf.substr(n1+1, n-n1-1);
h = str2int(height);
}
static int mode_to_prec[] = {
10,7,11,10,
10,7,11,11,
10,7,11,12,
10,7,9,16,
10,7,8,-1,
10,7,8,-1,
10,7,8,-1,
10,7,6,-1,
};
static int shapeindexhist[32], modehist[32], prechistone[16], prechisttwo[16], oneregion, tworegions;
static void stats(char block[ZOH::BLOCKSIZE])
{
char mode = block[0] & 0x1F; if ((mode & 0x3) == 0) mode = 0; if ((mode & 0x3) == 1) mode = 1; modehist[mode]++;
int prec = mode_to_prec[mode];
nvAssert (prec != -1);
if (!ZOH::isone(block))
{
tworegions++;
prechisttwo[prec]++;
int shapeindex = ((block[0] & 0xe0) >> 5) | ((block[1] & 0x3) << 3);
shapeindexhist[shapeindex]++;
}
else
{
oneregion++;
prechistone[prec]++;
}
}
static void printstats()
{
printf("\nPrecision histogram 10b to 16b one region: "); for (int i=10; i<=16; ++i) printf("%d,", prechistone[i]);
printf("\nPrecision histogram 6b to 11b two regions: "); for (int i=6; i<=11; ++i) printf("%d,", prechisttwo[i]);
printf("\nMode histogram: "); for (int i=0; i<32; ++i) printf("%d,", modehist[i]);
printf("\nShape index histogram: "); for (int i=0; i<32; ++i) printf("%d,", shapeindexhist[i]);
printf("\nOne region %5.2f%% Two regions %5.2f%%", 100.0*oneregion/float(oneregion+tworegions), 100.0*tworegions/float(oneregion+tworegions));
printf("\n");
}
void ZOH::decompress(string zohf, string outf)
{
Array2D<Rgba> pixels;
int w, h;
char block[ZOH::BLOCKSIZE];
extract(zohf, w, h);
FILE *zohfile = fopen(zohf.c_str(), "rb");
if (zohfile == NULL) throw "Unable to open .zoh file for read";
pixels.resizeErase(h, w);
// convert to tiles and decompress each tile
for (int y=0; y<h; y+=Tile::TILE_H)
{
int ysize = min(Tile::TILE_H, h-y);
for (int x=0; x<w; x+=Tile::TILE_W)
{
int xsize = min(Tile::TILE_W, w-x);
Tile t(xsize, ysize);
if (fread(block, sizeof(char), ZOH::BLOCKSIZE, zohfile) != ZOH::BLOCKSIZE)
throw "File error on read";
stats(block); // collect statistics
ZOH::decompress(block, t);
t.extract(pixels, x, y);
}
}
if (fclose(zohfile)) throw "Close failed on .zoh file";
Exr::writeRgba(outf, pixels, w, h);
#ifndef EXTERNAL_RELEASE
printstats(); // print statistics
#endif
}
*/

65
src/bc6h/zoh.h Normal file
View File

@ -0,0 +1,65 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
#pragma once
#ifndef _ZOH_H
#define _ZOH_H
#include "tile.h"
namespace ZOH {
// UNUSED ZOH MODES are 0x13, 0x17, 0x1b, 0x1f
static const int NREGIONS_TWO = 2;
static const int NREGIONS_ONE = 1;
static const int NCHANNELS = 3;
struct FltEndpts
{
nv::Vector3 A;
nv::Vector3 B;
};
struct IntEndpts
{
int A[NCHANNELS];
int B[NCHANNELS];
};
struct ComprEndpts
{
uint A[NCHANNELS];
uint B[NCHANNELS];
};
static const int BLOCKSIZE=16;
static const int BITSIZE=128;
void compress(const Tile &t, char *block);
void decompress(const char *block, Tile &t);
float compressone(const Tile &t, char *block);
float compresstwo(const Tile &t, char *block);
void decompressone(const char *block, Tile &t);
void decompresstwo(const char *block, Tile &t);
float refinetwo(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_TWO], char *block);
float roughtwo(const Tile &tile, int shape, FltEndpts endpts[NREGIONS_TWO]);
float refineone(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_ONE], char *block);
float roughone(const Tile &tile, int shape, FltEndpts endpts[NREGIONS_ONE]);
bool isone(const char *block);
}
#endif // _ZOH_H

324
src/bc6h/zoh_utils.cpp Normal file
View File

@ -0,0 +1,324 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
// Utility and common routines
#include "zoh_utils.h"
#include "nvmath/Vector.inl"
#include <math.h>
using namespace nv;
using namespace ZOH;
static const int denom7_weights_64[] = {0, 9, 18, 27, 37, 46, 55, 64}; // divided by 64
static const int denom15_weights_64[] = {0, 4, 9, 13, 17, 21, 26, 30, 34, 38, 43, 47, 51, 55, 60, 64}; // divided by 64
/*static*/ Format Utils::FORMAT;
int Utils::lerp(int a, int b, int i, int denom)
{
nvDebugCheck (denom == 3 || denom == 7 || denom == 15);
nvDebugCheck (i >= 0 && i <= denom);
int round = 32, shift = 6;
const int *weights;
switch(denom)
{
case 3: denom *= 5; i *= 5; // fall through to case 15
case 15: weights = denom15_weights_64; break;
case 7: weights = denom7_weights_64; break;
default: nvDebugCheck(0);
}
return (a*weights[denom-i] +b*weights[i] + round) >> shift;
}
Vector3 Utils::lerp(const Vector3& a, const Vector3 &b, int i, int denom)
{
nvDebugCheck (denom == 3 || denom == 7 || denom == 15);
nvDebugCheck (i >= 0 && i <= denom);
int shift = 6;
const int *weights;
switch(denom)
{
case 3: denom *= 5; i *= 5; // fall through to case 15
case 15: weights = denom15_weights_64; break;
case 7: weights = denom7_weights_64; break;
default: nvUnreachable();
}
// no need to round these as this is an exact division
return (a*float(weights[denom-i]) +b*float(weights[i])) / float(1 << shift);
}
/*
For unsigned f16, clamp the input to [0,F16MAX]. Thus u15.
For signed f16, clamp the input to [-F16MAX,F16MAX]. Thus s16.
The conversions proceed as follows:
unsigned f16: get bits. if high bit set, clamp to 0, else clamp to F16MAX.
signed f16: get bits. extract exp+mantissa and clamp to F16MAX. return -value if sign bit was set, else value
unsigned int: get bits. return as a positive value.
signed int. get bits. return as a value in -32768..32767.
The inverse conversions are just the inverse of the above.
*/
// clamp the 3 channels of the input vector to the allowable range based on FORMAT
// note that each channel is a float storing the allowable range as a bit pattern converted to float
// that is, for unsigned f16 say, we would clamp each channel to the range [0, F16MAX]
void Utils::clamp(Vector3 &v)
{
for (int i=0; i<3; ++i)
{
switch(Utils::FORMAT)
{
case UNSIGNED_F16:
if (v.component[i] < 0.0) v.component[i] = 0;
else if (v.component[i] > F16MAX) v.component[i] = F16MAX;
break;
case SIGNED_F16:
if (v.component[i] < -F16MAX) v.component[i] = -F16MAX;
else if (v.component[i] > F16MAX) v.component[i] = F16MAX;
break;
default:
nvUnreachable();
}
}
}
// convert a u16 value to s17 (represented as an int) based on the format expected
int Utils::ushort_to_format(unsigned short input)
{
int out, s;
// clamp to the valid range we are expecting
switch (Utils::FORMAT)
{
case UNSIGNED_F16:
if (input & F16S_MASK) out = 0;
else if (input > F16MAX) out = F16MAX;
else out = input;
break;
case SIGNED_F16:
s = input & F16S_MASK;
input &= F16EM_MASK;
if (input > F16MAX) out = F16MAX;
else out = input;
out = s ? -out : out;
break;
}
return out;
}
// convert a s17 value to u16 based on the format expected
unsigned short Utils::format_to_ushort(int input)
{
unsigned short out;
// clamp to the valid range we are expecting
switch (Utils::FORMAT)
{
case UNSIGNED_F16:
nvDebugCheck (input >= 0 && input <= F16MAX);
out = input;
break;
case SIGNED_F16:
nvDebugCheck (input >= -F16MAX && input <= F16MAX);
// convert to sign-magnitude
int s;
if (input < 0) { s = F16S_MASK; input = -input; }
else { s = 0; }
out = s | input;
break;
}
return out;
}
// quantize the input range into equal-sized bins
int Utils::quantize(float value, int prec)
{
int q, ivalue, s;
nvDebugCheck (prec > 1); // didn't bother to make it work for 1
value = (float)floor(value + 0.5);
int bias = (prec > 10) ? ((1<<(prec-1))-1) : 0; // bias precisions 11..16 to get a more accurate quantization
switch (Utils::FORMAT)
{
case UNSIGNED_F16:
nvDebugCheck (value >= 0 && value <= F16MAX);
ivalue = (int)value;
q = ((ivalue << prec) + bias) / (F16MAX+1);
nvDebugCheck (q >= 0 && q < (1 << prec));
break;
case SIGNED_F16:
nvDebugCheck (value >= -F16MAX && value <= F16MAX);
// convert to sign-magnitude
ivalue = (int)value;
if (ivalue < 0) { s = 1; ivalue = -ivalue; } else s = 0;
q = ((ivalue << (prec-1)) + bias) / (F16MAX+1);
if (s)
q = -q;
nvDebugCheck (q > -(1 << (prec-1)) && q < (1 << (prec-1)));
break;
}
return q;
}
int Utils::finish_unquantize(int q, int prec)
{
if (Utils::FORMAT == UNSIGNED_F16)
return (q * 31) >> 6; // scale the magnitude by 31/64
else if (Utils::FORMAT == SIGNED_F16)
return (q < 0) ? -(((-q) * 31) >> 5) : (q * 31) >> 5; // scale the magnitude by 31/32
else
return q;
}
// unquantize each bin to midpoint of original bin range, except
// for the end bins which we push to an endpoint of the bin range.
// we do this to ensure we can represent all possible original values.
// the asymmetric end bins do not affect PSNR for the test images.
//
// code this function assuming an arbitrary bit pattern as the encoded block
int Utils::unquantize(int q, int prec)
{
int unq, s;
nvDebugCheck (prec > 1); // not implemented for prec 1
switch (Utils::FORMAT)
{
// modify this case to move the multiplication by 31 after interpolation.
// Need to use finish_unquantize.
// since we have 16 bits available, let's unquantize this to 16 bits unsigned
// thus the scale factor is [0-7c00)/[0-10000) = 31/64
case UNSIGNED_F16:
if (prec >= 15)
unq = q;
else if (q == 0)
unq = 0;
else if (q == ((1<<prec)-1))
unq = U16MAX;
else
unq = (q * (U16MAX+1) + (U16MAX+1)/2) >> prec;
break;
// here, let's stick with S16 (no apparent quality benefit from going to S17)
// range is (-7c00..7c00)/(-8000..8000) = 31/32
case SIGNED_F16:
// don't remove this test even though it appears equivalent to the code below
// as it isn't -- the code below can overflow for prec = 16
if (prec >= 16)
unq = q;
else
{
if (q < 0) { s = 1; q = -q; } else s = 0;
if (q == 0)
unq = 0;
else if (q >= ((1<<(prec-1))-1))
unq = s ? -S16MAX : S16MAX;
else
{
unq = (q * (S16MAX+1) + (S16MAX+1)/2) >> (prec-1);
if (s)
unq = -unq;
}
}
break;
}
return unq;
}
// pick a norm!
#define NORM_EUCLIDEAN 1
float Utils::norm(const Vector3 &a, const Vector3 &b)
{
#ifdef NORM_EUCLIDEAN
return lengthSquared(a - b);
#endif
#ifdef NORM_ABS
Vector3 err = a - b;
return fabs(err.x) + fabs(err.y) + fabs(err.z);
#endif
}
// parse <name>[<start>{:<end>}]{,}
// the pointer starts here ^
// name is 1 or 2 chars and matches field names. start and end are decimal numbers
void Utils::parse(const char *encoding, int &ptr, Field &field, int &endbit, int &len)
{
if (ptr <= 0) return;
--ptr;
if (encoding[ptr] == ',') --ptr;
nvDebugCheck (encoding[ptr] == ']');
--ptr;
endbit = 0;
int scale = 1;
while (encoding[ptr] != ':' && encoding[ptr] != '[')
{
nvDebugCheck(encoding[ptr] >= '0' && encoding[ptr] <= '9');
endbit += (encoding[ptr--] - '0') * scale;
scale *= 10;
}
int startbit = 0; scale = 1;
if (encoding[ptr] == '[')
startbit = endbit;
else
{
ptr--;
while (encoding[ptr] != '[')
{
nvDebugCheck(encoding[ptr] >= '0' && encoding[ptr] <= '9');
startbit += (encoding[ptr--] - '0') * scale;
scale *= 10;
}
}
len = startbit - endbit + 1; // startbit>=endbit note
--ptr;
if (encoding[ptr] == 'm') field = FIELD_M;
else if (encoding[ptr] == 'd') field = FIELD_D;
else {
// it's wxyz
nvDebugCheck (encoding[ptr] >= 'w' && encoding[ptr] <= 'z');
int foo = encoding[ptr--] - 'w';
// now it is r g or b
if (encoding[ptr] == 'r') foo += 10;
else if (encoding[ptr] == 'g') foo += 20;
else if (encoding[ptr] == 'b') foo += 30;
else nvDebugCheck(0);
field = (Field) foo;
}
}

73
src/bc6h/zoh_utils.h Normal file
View File

@ -0,0 +1,73 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
// utility class holding common routines
#pragma once
#ifndef _ZOH_UTILS_H
#define _ZOH_UTILS_H
#include "nvmath/Vector.h"
namespace ZOH {
inline int SIGN_EXTEND(int x, int nb) { return ((((signed(x))&(1<<((nb)-1)))?((~0)<<(nb)):0)|(signed(x))); }
enum Field {
FIELD_M = 1, // mode
FIELD_D = 2, // distribution/shape
FIELD_RW = 10+0, FIELD_RX = 10+1, FIELD_RY = 10+2, FIELD_RZ = 10+3, // red channel endpoints or deltas
FIELD_GW = 20+0, FIELD_GX = 20+1, FIELD_GY = 20+2, FIELD_GZ = 20+3, // green channel endpoints or deltas
FIELD_BW = 30+0, FIELD_BX = 30+1, FIELD_BY = 30+2, FIELD_BZ = 30+3, // blue channel endpoints or deltas
};
// some constants
static const int F16S_MASK = 0x8000; // f16 sign mask
static const int F16EM_MASK = 0x7fff; // f16 exp & mantissa mask
static const int U16MAX = 0xffff;
static const int S16MIN = -0x8000;
static const int S16MAX = 0x7fff;
static const int INT16_MASK = 0xffff;
static const int F16MAX = 0x7bff; // MAXFLT bit pattern for halfs
enum Format { UNSIGNED_F16, SIGNED_F16 };
class Utils
{
public:
static Format FORMAT; // this is a global -- we're either handling unsigned or unsigned half values
// error metrics
static float norm(const nv::Vector3 &a, const nv::Vector3 &b);
static float mpsnr_norm(const nv::Vector3 &a, int exposure, const nv::Vector3 &b);
// conversion & clamp
static int ushort_to_format(unsigned short input);
static unsigned short format_to_ushort(int input);
// clamp to format
static void clamp(nv::Vector3 &v);
// quantization and unquantization
static int finish_unquantize(int q, int prec);
static int unquantize(int q, int prec);
static int quantize(float value, int prec);
static void parse(const char *encoding, int &ptr, Field & field, int &endbit, int &len);
// lerping
static int lerp(int a, int b, int i, int denom);
static nv::Vector3 lerp(const nv::Vector3 & a, const nv::Vector3 & b, int i, int denom);
};
}
#endif // _ZOH_UTILS_H

799
src/bc6h/zohone.cpp Normal file
View File

@ -0,0 +1,799 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
// one region zoh compress/decompress code
// Thanks to Jacob Munkberg (jacob@cs.lth.se) for the shortcut of using SVD to do the equivalent of principal components analysis
#include "bits.h"
#include "tile.h"
#include "zoh.h"
#include "zoh_utils.h"
#include "nvmath/Vector.inl"
#include "nvmath/Fitting.h"
#include <string.h> // strlen
#include <float.h> // FLT_MAX
using namespace nv;
using namespace ZOH;
#define NINDICES 16
#define INDEXBITS 4
#define HIGH_INDEXBIT (1<<(INDEXBITS-1))
#define DENOM (NINDICES-1)
#define NSHAPES 1
static const int shapes[NSHAPES] =
{
0x0000
}; // only 1 shape
#define REGION(x,y,shapeindex) ((shapes[shapeindex]&(1<<(15-(x)-4*(y))))!=0)
#define POS_TO_X(pos) ((pos)&3)
#define POS_TO_Y(pos) (((pos)>>2)&3)
#define NDELTA 2
struct Chanpat
{
int prec[NDELTA]; // precision pattern for one channel
};
struct Pattern
{
Chanpat chan[NCHANNELS];// allow different bit patterns per channel -- but we still want constant precision per channel
int transformed; // if 0, deltas are unsigned and no transform; otherwise, signed and transformed
int mode; // associated mode value
int modebits; // number of mode bits
const char *encoding; // verilog description of encoding for this mode
};
#define MAXMODEBITS 5
#define MAXMODES (1<<MAXMODEBITS)
#define NPATTERNS 4
static const Pattern patterns[NPATTERNS] =
{
16,4, 16,4, 16,4, 1, 0x0f, 5, "bw[10],bw[11],bw[12],bw[13],bw[14],bw[15],bx[3:0],gw[10],gw[11],gw[12],gw[13],gw[14],gw[15],gx[3:0],rw[10],rw[11],rw[12],rw[13],rw[14],rw[15],rx[3:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
12,8, 12,8, 12,8, 1, 0x0b, 5, "bw[10],bw[11],bx[7:0],gw[10],gw[11],gx[7:0],rw[10],rw[11],rx[7:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
11,9, 11,9, 11,9, 1, 0x07, 5, "bw[10],bx[8:0],gw[10],gx[8:0],rw[10],rx[8:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
10,10, 10,10, 10,10, 0, 0x03, 5, "bx[9:0],gx[9:0],rx[9:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
};
// mapping of mode to the corresponding index in pattern
static const int mode_to_pat[MAXMODES] = {
-1,-1,-1,
3, // 0x03
-1,-1,-1,
2, // 0x07
-1,-1,-1,
1, // 0x0b
-1,-1,-1,
0, // 0x0f
-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,-1,
};
#define R_0(ep) (ep)[0].A[i]
#define R_1(ep) (ep)[0].B[i]
#define MASK(n) ((1<<(n))-1)
// compress endpoints
static void compress_endpts(const IntEndpts in[NREGIONS_ONE], ComprEndpts out[NREGIONS_ONE], const Pattern &p)
{
if (p.transformed)
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = R_0(in) & MASK(p.chan[i].prec[0]);
R_1(out) = (R_1(in) - R_0(in)) & MASK(p.chan[i].prec[1]);
}
}
else
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = R_0(in) & MASK(p.chan[i].prec[0]);
R_1(out) = R_1(in) & MASK(p.chan[i].prec[1]);
}
}
}
// decompress endpoints
static void decompress_endpts(const ComprEndpts in[NREGIONS_ONE], IntEndpts out[NREGIONS_ONE], const Pattern &p)
{
bool issigned = Utils::FORMAT == SIGNED_F16;
if (p.transformed)
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = issigned ? SIGN_EXTEND(R_0(in),p.chan[i].prec[0]) : R_0(in);
int t;
t = SIGN_EXTEND(R_1(in), p.chan[i].prec[1]);
t = (t + R_0(in)) & MASK(p.chan[i].prec[0]);
R_1(out) = issigned ? SIGN_EXTEND(t,p.chan[i].prec[0]) : t;
}
}
else
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = issigned ? SIGN_EXTEND(R_0(in),p.chan[i].prec[0]) : R_0(in);
R_1(out) = issigned ? SIGN_EXTEND(R_1(in),p.chan[i].prec[1]) : R_1(in);
}
}
}
static void quantize_endpts(const FltEndpts endpts[NREGIONS_ONE], int prec, IntEndpts q_endpts[NREGIONS_ONE])
{
for (int region = 0; region < NREGIONS_ONE; ++region)
{
q_endpts[region].A[0] = Utils::quantize(endpts[region].A.x, prec);
q_endpts[region].A[1] = Utils::quantize(endpts[region].A.y, prec);
q_endpts[region].A[2] = Utils::quantize(endpts[region].A.z, prec);
q_endpts[region].B[0] = Utils::quantize(endpts[region].B.x, prec);
q_endpts[region].B[1] = Utils::quantize(endpts[region].B.y, prec);
q_endpts[region].B[2] = Utils::quantize(endpts[region].B.z, prec);
}
}
// swap endpoints as needed to ensure that the indices at index_one and index_one have a 0 high-order bit
// index_one is 0 at x=0 y=0 and 15 at x=3 y=3 so y = (index >> 2) & 3 and x = index & 3
static void swap_indices(IntEndpts endpts[NREGIONS_ONE], int indices[Tile::TILE_H][Tile::TILE_W], int shapeindex)
{
int index_positions[NREGIONS_ONE];
index_positions[0] = 0; // since WLOG we have the high bit of the shapes at 0
for (int region = 0; region < NREGIONS_ONE; ++region)
{
int x = index_positions[region] & 3;
int y = (index_positions[region] >> 2) & 3;
nvDebugCheck(REGION(x,y,shapeindex) == region); // double check the table
if (indices[y][x] & HIGH_INDEXBIT)
{
// high bit is set, swap the endpts and indices for this region
int t;
for (int i=0; i<NCHANNELS; ++i) { t = endpts[region].A[i]; endpts[region].A[i] = endpts[region].B[i]; endpts[region].B[i] = t; }
for (int y = 0; y < Tile::TILE_H; y++)
for (int x = 0; x < Tile::TILE_W; x++)
if (REGION(x,y,shapeindex) == region)
indices[y][x] = NINDICES - 1 - indices[y][x];
}
}
}
// endpoints fit only if the compression was lossless
static bool endpts_fit(const IntEndpts orig[NREGIONS_ONE], const ComprEndpts compressed[NREGIONS_ONE], const Pattern &p)
{
IntEndpts uncompressed[NREGIONS_ONE];
decompress_endpts(compressed, uncompressed, p);
for (int j=0; j<NREGIONS_ONE; ++j)
for (int i=0; i<NCHANNELS; ++i)
{
if (orig[j].A[i] != uncompressed[j].A[i]) return false;
if (orig[j].B[i] != uncompressed[j].B[i]) return false;
}
return true;
}
static void write_header(const ComprEndpts endpts[NREGIONS_ONE], const Pattern &p, Bits &out)
{
// interpret the verilog backwards and process it
int m = p.mode;
int rw = endpts[0].A[0], rx = endpts[0].B[0];
int gw = endpts[0].A[1], gx = endpts[0].B[1];
int bw = endpts[0].A[2], bx = endpts[0].B[2];
int ptr = int(strlen(p.encoding));
while (ptr)
{
Field field;
int endbit, len;
// !!!UNDONE: get rid of string parsing!!!
Utils::parse(p.encoding, ptr, field, endbit, len);
switch(field)
{
case FIELD_M: out.write( m >> endbit, len); break;
case FIELD_RW: out.write(rw >> endbit, len); break;
case FIELD_RX: out.write(rx >> endbit, len); break;
case FIELD_GW: out.write(gw >> endbit, len); break;
case FIELD_GX: out.write(gx >> endbit, len); break;
case FIELD_BW: out.write(bw >> endbit, len); break;
case FIELD_BX: out.write(bx >> endbit, len); break;
case FIELD_D:
case FIELD_RY:
case FIELD_RZ:
case FIELD_GY:
case FIELD_GZ:
case FIELD_BY:
case FIELD_BZ:
default: nvUnreachable();
}
}
}
static void read_header(Bits &in, ComprEndpts endpts[NREGIONS_ONE], Pattern &p)
{
// reading isn't quite symmetric with writing -- we don't know the encoding until we decode the mode
int mode = in.read(2);
if (mode != 0x00 && mode != 0x01)
mode = (in.read(3) << 2) | mode;
int pat_index = mode_to_pat[mode];
nvDebugCheck (pat_index >= 0 && pat_index < NPATTERNS);
nvDebugCheck (in.getptr() == patterns[pat_index].modebits);
p = patterns[pat_index];
int d;
int rw, rx;
int gw, gx;
int bw, bx;
d = 0;
rw = rx = 0;
gw = gx = 0;
bw = bx = 0;
int ptr = int(strlen(p.encoding));
while (ptr)
{
Field field;
int endbit, len;
// !!!UNDONE: get rid of string parsing!!!
Utils::parse(p.encoding, ptr, field, endbit, len);
switch(field)
{
case FIELD_M: break; // already processed so ignore
case FIELD_RW: rw |= in.read(len) << endbit; break;
case FIELD_RX: rx |= in.read(len) << endbit; break;
case FIELD_GW: gw |= in.read(len) << endbit; break;
case FIELD_GX: gx |= in.read(len) << endbit; break;
case FIELD_BW: bw |= in.read(len) << endbit; break;
case FIELD_BX: bx |= in.read(len) << endbit; break;
case FIELD_D:
case FIELD_RY:
case FIELD_RZ:
case FIELD_GY:
case FIELD_GZ:
case FIELD_BY:
case FIELD_BZ:
default: nvUnreachable();
}
}
nvDebugCheck (in.getptr() == 128 - 63);
endpts[0].A[0] = rw; endpts[0].B[0] = rx;
endpts[0].A[1] = gw; endpts[0].B[1] = gx;
endpts[0].A[2] = bw; endpts[0].B[2] = bx;
}
// compress index 0
static void write_indices(const int indices[Tile::TILE_H][Tile::TILE_W], int shapeindex, Bits &out)
{
for (int pos = 0; pos < Tile::TILE_TOTAL; ++pos)
{
int x = POS_TO_X(pos);
int y = POS_TO_Y(pos);
out.write(indices[y][x], INDEXBITS - ((pos == 0) ? 1 : 0));
}
}
static void emit_block(const ComprEndpts endpts[NREGIONS_ONE], int shapeindex, const Pattern &p, const int indices[Tile::TILE_H][Tile::TILE_W], char *block)
{
Bits out(block, ZOH::BITSIZE);
write_header(endpts, p, out);
write_indices(indices, shapeindex, out);
nvDebugCheck(out.getptr() == ZOH::BITSIZE);
}
static void generate_palette_quantized(const IntEndpts &endpts, int prec, Vector3 palette[NINDICES])
{
// scale endpoints
int a, b; // really need a IntVector3...
a = Utils::unquantize(endpts.A[0], prec);
b = Utils::unquantize(endpts.B[0], prec);
// interpolate
for (int i = 0; i < NINDICES; ++i)
palette[i].x = float(Utils::finish_unquantize(Utils::lerp(a, b, i, DENOM), prec));
a = Utils::unquantize(endpts.A[1], prec);
b = Utils::unquantize(endpts.B[1], prec);
// interpolate
for (int i = 0; i < NINDICES; ++i)
palette[i].y = float(Utils::finish_unquantize(Utils::lerp(a, b, i, DENOM), prec));
a = Utils::unquantize(endpts.A[2], prec);
b = Utils::unquantize(endpts.B[2], prec);
// interpolate
for (int i = 0; i < NINDICES; ++i)
palette[i].z = float(Utils::finish_unquantize(Utils::lerp(a, b, i, DENOM), prec));
}
// position 0 was compressed
static void read_indices(Bits &in, int shapeindex, int indices[Tile::TILE_H][Tile::TILE_W])
{
for (int pos = 0; pos < Tile::TILE_TOTAL; ++pos)
{
int x = POS_TO_X(pos);
int y = POS_TO_Y(pos);
indices[y][x]= in.read(INDEXBITS - ((pos == 0) ? 1 : 0));
}
}
void ZOH::decompressone(const char *block, Tile &t)
{
Bits in(block, ZOH::BITSIZE);
Pattern p;
IntEndpts endpts[NREGIONS_ONE];
ComprEndpts compr_endpts[NREGIONS_ONE];
read_header(in, compr_endpts, p);
int shapeindex = 0; // only one shape
decompress_endpts(compr_endpts, endpts, p);
Vector3 palette[NREGIONS_ONE][NINDICES];
for (int r = 0; r < NREGIONS_ONE; ++r)
generate_palette_quantized(endpts[r], p.chan[0].prec[0], &palette[r][0]);
// read indices
int indices[Tile::TILE_H][Tile::TILE_W];
read_indices(in, shapeindex, indices);
nvDebugCheck(in.getptr() == ZOH::BITSIZE);
// lookup
for (int y = 0; y < Tile::TILE_H; y++)
for (int x = 0; x < Tile::TILE_W; x++)
t.data[y][x] = palette[REGION(x,y,shapeindex)][indices[y][x]];
}
// given a collection of colors and quantized endpoints, generate a palette, choose best entries, and return a single toterr
static float map_colors(const Vector3 colors[], const float importance[], int np, const IntEndpts &endpts, int prec)
{
Vector3 palette[NINDICES];
float toterr = 0;
Vector3 err;
generate_palette_quantized(endpts, prec, palette);
for (int i = 0; i < np; ++i)
{
float err, besterr;
besterr = Utils::norm(colors[i], palette[0]) * importance[i];
for (int j = 1; j < NINDICES && besterr > 0; ++j)
{
err = Utils::norm(colors[i], palette[j]) * importance[i];
if (err > besterr) // error increased, so we're done searching
break;
if (err < besterr)
besterr = err;
}
toterr += besterr;
}
return toterr;
}
// assign indices given a tile, shape, and quantized endpoints, return toterr for each region
static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NREGIONS_ONE], int prec,
int indices[Tile::TILE_H][Tile::TILE_W], float toterr[NREGIONS_ONE])
{
// build list of possibles
Vector3 palette[NREGIONS_ONE][NINDICES];
for (int region = 0; region < NREGIONS_ONE; ++region)
{
generate_palette_quantized(endpts[region], prec, &palette[region][0]);
toterr[region] = 0;
}
Vector3 err;
for (int y = 0; y < tile.size_y; y++)
for (int x = 0; x < tile.size_x; x++)
{
int region = REGION(x,y,shapeindex);
float err, besterr;
besterr = Utils::norm(tile.data[y][x], palette[region][0]);
indices[y][x] = 0;
for (int i = 1; i < NINDICES && besterr > 0; ++i)
{
err = Utils::norm(tile.data[y][x], palette[region][i]);
if (err > besterr) // error increased, so we're done searching
break;
if (err < besterr)
{
besterr = err;
indices[y][x] = i;
}
}
toterr[region] += besterr;
}
}
static float perturb_one(const Vector3 colors[], const float importance[], int np, int ch, int prec, const IntEndpts &old_endpts, IntEndpts &new_endpts,
float old_err, int do_b)
{
// we have the old endpoints: old_endpts
// we have the perturbed endpoints: new_endpts
// we have the temporary endpoints: temp_endpts
IntEndpts temp_endpts;
float min_err = old_err; // start with the best current error
int beststep;
// copy real endpoints so we can perturb them
for (int i=0; i<NCHANNELS; ++i) { temp_endpts.A[i] = new_endpts.A[i] = old_endpts.A[i]; temp_endpts.B[i] = new_endpts.B[i] = old_endpts.B[i]; }
// do a logarithmic search for the best error for this endpoint (which)
for (int step = 1 << (prec-1); step; step >>= 1)
{
bool improved = false;
for (int sign = -1; sign <= 1; sign += 2)
{
if (do_b == 0)
{
temp_endpts.A[ch] = new_endpts.A[ch] + sign * step;
if (temp_endpts.A[ch] < 0 || temp_endpts.A[ch] >= (1 << prec))
continue;
}
else
{
temp_endpts.B[ch] = new_endpts.B[ch] + sign * step;
if (temp_endpts.B[ch] < 0 || temp_endpts.B[ch] >= (1 << prec))
continue;
}
float err = map_colors(colors, importance, np, temp_endpts, prec);
if (err < min_err)
{
improved = true;
min_err = err;
beststep = sign * step;
}
}
// if this was an improvement, move the endpoint and continue search from there
if (improved)
{
if (do_b == 0)
new_endpts.A[ch] += beststep;
else
new_endpts.B[ch] += beststep;
}
}
return min_err;
}
static void optimize_one(const Vector3 colors[], const float importance[], int np, float orig_err, const IntEndpts &orig_endpts, int prec, IntEndpts &opt_endpts)
{
float opt_err = orig_err;
for (int ch = 0; ch < NCHANNELS; ++ch)
{
opt_endpts.A[ch] = orig_endpts.A[ch];
opt_endpts.B[ch] = orig_endpts.B[ch];
}
/*
err0 = perturb(rgb0, delta0)
err1 = perturb(rgb1, delta1)
if (err0 < err1)
if (err0 >= initial_error) break
rgb0 += delta0
next = 1
else
if (err1 >= initial_error) break
rgb1 += delta1
next = 0
initial_err = map()
for (;;)
err = perturb(next ? rgb1:rgb0, delta)
if (err >= initial_err) break
next? rgb1 : rgb0 += delta
initial_err = err
*/
IntEndpts new_a, new_b;
IntEndpts new_endpt;
int do_b;
// now optimize each channel separately
for (int ch = 0; ch < NCHANNELS; ++ch)
{
// figure out which endpoint when perturbed gives the most improvement and start there
// if we just alternate, we can easily end up in a local minima
float err0 = perturb_one(colors, importance, np, ch, prec, opt_endpts, new_a, opt_err, 0); // perturb endpt A
float err1 = perturb_one(colors, importance, np, ch, prec, opt_endpts, new_b, opt_err, 1); // perturb endpt B
if (err0 < err1)
{
if (err0 >= opt_err)
continue;
opt_endpts.A[ch] = new_a.A[ch];
opt_err = err0;
do_b = 1; // do B next
}
else
{
if (err1 >= opt_err)
continue;
opt_endpts.B[ch] = new_b.B[ch];
opt_err = err1;
do_b = 0; // do A next
}
// now alternate endpoints and keep trying until there is no improvement
for (;;)
{
float err = perturb_one(colors, importance, np, ch, prec, opt_endpts, new_endpt, opt_err, do_b);
if (err >= opt_err)
break;
if (do_b == 0)
opt_endpts.A[ch] = new_endpt.A[ch];
else
opt_endpts.B[ch] = new_endpt.B[ch];
opt_err = err;
do_b = 1 - do_b; // now move the other endpoint
}
}
}
static void optimize_endpts(const Tile &tile, int shapeindex, const float orig_err[NREGIONS_ONE],
const IntEndpts orig_endpts[NREGIONS_ONE], int prec, IntEndpts opt_endpts[NREGIONS_ONE])
{
Vector3 pixels[Tile::TILE_TOTAL];
float importance[Tile::TILE_TOTAL];
float err = 0;
for (int region=0; region<NREGIONS_ONE; ++region)
{
// collect the pixels in the region
int np = 0;
for (int y = 0; y < tile.size_y; y++) {
for (int x = 0; x < tile.size_x; x++) {
if (REGION(x, y, shapeindex) == region) {
pixels[np] = tile.data[y][x];
importance[np] = tile.importance_map[y][x];
++np;
}
}
}
optimize_one(pixels, importance, np, orig_err[region], orig_endpts[region], prec, opt_endpts[region]);
}
}
/* optimization algorithm
for each pattern
convert endpoints using pattern precision
assign indices and get initial error
compress indices (and possibly reorder endpoints)
transform endpoints
if transformed endpoints fit pattern
get original endpoints back
optimize endpoints, get new endpoints, new indices, and new error // new error will almost always be better
compress new indices
transform new endpoints
if new endpoints fit pattern AND if error is improved
emit compressed block with new data
else
emit compressed block with original data // to try to preserve maximum endpoint precision
*/
float ZOH::refineone(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_ONE], char *block)
{
float orig_err[NREGIONS_ONE], opt_err[NREGIONS_ONE], orig_toterr, opt_toterr;
IntEndpts orig_endpts[NREGIONS_ONE], opt_endpts[NREGIONS_ONE];
ComprEndpts compr_orig[NREGIONS_ONE], compr_opt[NREGIONS_ONE];
int orig_indices[Tile::TILE_H][Tile::TILE_W], opt_indices[Tile::TILE_H][Tile::TILE_W];
for (int sp = 0; sp < NPATTERNS; ++sp)
{
// precisions for all channels need to be the same
for (int i=1; i<NCHANNELS; ++i) nvDebugCheck (patterns[sp].chan[0].prec[0] == patterns[sp].chan[i].prec[0]);
quantize_endpts(endpts, patterns[sp].chan[0].prec[0], orig_endpts);
assign_indices(tile, shapeindex_best, orig_endpts, patterns[sp].chan[0].prec[0], orig_indices, orig_err);
swap_indices(orig_endpts, orig_indices, shapeindex_best);
compress_endpts(orig_endpts, compr_orig, patterns[sp]);
if (endpts_fit(orig_endpts, compr_orig, patterns[sp]))
{
optimize_endpts(tile, shapeindex_best, orig_err, orig_endpts, patterns[sp].chan[0].prec[0], opt_endpts);
assign_indices(tile, shapeindex_best, opt_endpts, patterns[sp].chan[0].prec[0], opt_indices, opt_err);
swap_indices(opt_endpts, opt_indices, shapeindex_best);
compress_endpts(opt_endpts, compr_opt, patterns[sp]);
orig_toterr = opt_toterr = 0;
for (int i=0; i < NREGIONS_ONE; ++i) { orig_toterr += orig_err[i]; opt_toterr += opt_err[i]; }
if (endpts_fit(opt_endpts, compr_opt, patterns[sp]) && opt_toterr < orig_toterr)
{
emit_block(compr_opt, shapeindex_best, patterns[sp], opt_indices, block);
return opt_toterr;
}
else
{
// either it stopped fitting when we optimized it, or there was no improvement
// so go back to the unoptimized endpoints which we know will fit
emit_block(compr_orig, shapeindex_best, patterns[sp], orig_indices, block);
return orig_toterr;
}
}
}
nvAssert (false); // "No candidate found, should never happen (refineone.)";
return FLT_MAX;
}
static void generate_palette_unquantized(const FltEndpts endpts[NREGIONS_ONE], Vector3 palette[NREGIONS_ONE][NINDICES])
{
for (int region = 0; region < NREGIONS_ONE; ++region)
for (int i = 0; i < NINDICES; ++i)
palette[region][i] = Utils::lerp(endpts[region].A, endpts[region].B, i, DENOM);
}
// generate a palette from unquantized endpoints, then pick best palette color for all pixels in each region, return toterr for all regions combined
static float map_colors(const Tile &tile, int shapeindex, const FltEndpts endpts[NREGIONS_ONE])
{
// build list of possibles
Vector3 palette[NREGIONS_ONE][NINDICES];
generate_palette_unquantized(endpts, palette);
float toterr = 0;
Vector3 err;
for (int y = 0; y < tile.size_y; y++)
for (int x = 0; x < tile.size_x; x++)
{
int region = REGION(x,y,shapeindex);
float err, besterr;
besterr = Utils::norm(tile.data[y][x], palette[region][0]) * tile.importance_map[y][x];
for (int i = 1; i < NINDICES && besterr > 0; ++i)
{
err = Utils::norm(tile.data[y][x], palette[region][i]) * tile.importance_map[y][x];
if (err > besterr) // error increased, so we're done searching
break;
if (err < besterr)
besterr = err;
}
toterr += besterr;
}
return toterr;
}
float ZOH::roughone(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS_ONE])
{
for (int region=0; region<NREGIONS_ONE; ++region)
{
int np = 0;
Vector3 colors[Tile::TILE_TOTAL];
Vector3 mean(0,0,0);
for (int y = 0; y < tile.size_y; y++) {
for (int x = 0; x < tile.size_x; x++) {
if (REGION(x,y,shapeindex) == region)
{
colors[np] = tile.data[y][x];
mean += tile.data[y][x];
++np;
}
}
}
// handle simple cases
if (np == 0)
{
Vector3 zero(0,0,0);
endpts[region].A = zero;
endpts[region].B = zero;
continue;
}
else if (np == 1)
{
endpts[region].A = colors[0];
endpts[region].B = colors[0];
continue;
}
else if (np == 2)
{
endpts[region].A = colors[0];
endpts[region].B = colors[1];
continue;
}
mean /= float(np);
Vector3 direction = Fit::computePrincipalComponent_EigenSolver(np, colors);
// project each pixel value along the principal direction
float minp = FLT_MAX, maxp = -FLT_MAX;
for (int i = 0; i < np; i++)
{
float dp = dot(colors[i]-mean, direction);
if (dp < minp) minp = dp;
if (dp > maxp) maxp = dp;
}
// choose as endpoints 2 points along the principal direction that span the projections of all of the pixel values
endpts[region].A = mean + minp*direction;
endpts[region].B = mean + maxp*direction;
// clamp endpoints
// the argument for clamping is that the actual endpoints need to be clamped and thus we need to choose the best
// shape based on endpoints being clamped
Utils::clamp(endpts[region].A);
Utils::clamp(endpts[region].B);
}
return map_colors(tile, shapeindex, endpts);
}
float ZOH::compressone(const Tile &t, char *block)
{
int shapeindex_best = 0;
FltEndpts endptsbest[NREGIONS_ONE], tempendpts[NREGIONS_ONE];
float msebest = FLT_MAX;
/*
collect the mse values that are within 5% of the best values
optimize each one and choose the best
*/
// hack for now -- just use the best value WORK
for (int i=0; i<NSHAPES && msebest>0.0; ++i)
{
float mse = roughone(t, i, tempendpts);
if (mse < msebest)
{
msebest = mse;
shapeindex_best = i;
memcpy(endptsbest, tempendpts, sizeof(endptsbest));
}
}
return refineone(t, shapeindex_best, endptsbest, block);
}

883
src/bc6h/zohtwo.cpp Normal file
View File

@ -0,0 +1,883 @@
/*
Copyright 2007 nVidia, Inc.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License.
You may obtain a copy of the License at http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and limitations under the License.
*/
// two regions zoh compress/decompress code
// Thanks to Jacob Munkberg (jacob@cs.lth.se) for the shortcut of using SVD to do the equivalent of principal components analysis
/* optimization algorithm
get initial float endpoints
convert endpoints using 16 bit precision, transform, and get bit delta. choose likely endpoint compression candidates.
note that there will be 1 or 2 candidates; 2 will be chosen when the delta values are close to the max possible.
for each EC candidate in order from max precision to smaller precision
convert endpoints using the appropriate precision.
optimize the endpoints and minimize square error. save the error and index assignments. apply index compression as well.
(thus the endpoints and indices are in final form.)
transform and get bit delta.
if the bit delta fits, exit
if we ended up with no candidates somehow, choose the tail set of EC candidates and retry. this should happen hardly ever.
add a state variable to nvDebugCheck we only do this once.
convert to bit stream.
return the error.
Global optimization
order all tiles based on their errors
do something special for high-error tiles
the goal here is to try to avoid tiling artifacts. but I think this is a research problem. let's just generate an error image...
display an image that shows partitioning and precision selected for each tile
*/
#include "bits.h"
#include "tile.h"
#include "zoh.h"
#include "zoh_utils.h"
#include "nvmath/Fitting.h"
#include "nvmath/Vector.inl"
#include <string.h> // strlen
#include <float.h> // FLT_MAX
using namespace nv;
using namespace ZOH;
#define NINDICES 8
#define INDEXBITS 3
#define HIGH_INDEXBIT (1<<(INDEXBITS-1))
#define DENOM (NINDICES-1)
// WORK: determine optimal traversal pattern to search for best shape -- what does the error curve look like?
// i.e. can we search shapes in a particular order so we can see the global error minima easily and
// stop without having to touch all shapes?
#include "shapes_two.h"
// use only the first 32 available shapes
#undef NSHAPES
#undef SHAPEBITS
#define NSHAPES 32
#define SHAPEBITS 5
#define POS_TO_X(pos) ((pos)&3)
#define POS_TO_Y(pos) (((pos)>>2)&3)
#define NDELTA 4
struct Chanpat
{
int prec[NDELTA]; // precision pattern for one channel
};
struct Pattern
{
Chanpat chan[NCHANNELS]; // allow different bit patterns per channel -- but we still want constant precision per channel
int transformed; // if 0, deltas are unsigned and no transform; otherwise, signed and transformed
int mode; // associated mode value
int modebits; // number of mode bits
const char *encoding; // verilog description of encoding for this mode
};
#define MAXMODEBITS 5
#define MAXMODES (1<<MAXMODEBITS)
#define NPATTERNS 10
static const Pattern patterns[NPATTERNS] =
{
11,5,5,5, 11,4,4,4, 11,4,4,4, 1, 0x02, 5, "d[4:0],bz[3],rz[4:0],bz[2],ry[4:0],by[3:0],bz[1],bw[10],bx[3:0],gz[3:0],bz[0],gw[10],gx[3:0],gy[3:0],rw[10],rx[4:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
11,4,4,4, 11,5,5,5, 11,4,4,4, 1, 0x06, 5, "d[4:0],bz[3],gy[4],rz[3:0],bz[2],bz[0],ry[3:0],by[3:0],bz[1],bw[10],bx[3:0],gz[3:0],gw[10],gx[4:0],gy[3:0],gz[4],rw[10],rx[3:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
11,4,4,4, 11,4,4,4, 11,5,5,5, 1, 0x0a, 5, "d[4:0],bz[3],bz[4],rz[3:0],bz[2:1],ry[3:0],by[3:0],bw[10],bx[4:0],gz[3:0],bz[0],gw[10],gx[3:0],gy[3:0],by[4],rw[10],rx[3:0],bw[9:0],gw[9:0],rw[9:0],m[4:0]",
10,5,5,5, 10,5,5,5, 10,5,5,5, 1, 0x00, 2, "d[4:0],bz[3],rz[4:0],bz[2],ry[4:0],by[3:0],bz[1],bx[4:0],gz[3:0],bz[0],gx[4:0],gy[3:0],gz[4],rx[4:0],bw[9:0],gw[9:0],rw[9:0],bz[4],by[4],gy[4],m[1:0]",
9,5,5,5, 9,5,5,5, 9,5,5,5, 1, 0x0e, 5, "d[4:0],bz[3],rz[4:0],bz[2],ry[4:0],by[3:0],bz[1],bx[4:0],gz[3:0],bz[0],gx[4:0],gy[3:0],gz[4],rx[4:0],bz[4],bw[8:0],gy[4],gw[8:0],by[4],rw[8:0],m[4:0]",
8,6,6,6, 8,5,5,5, 8,5,5,5, 1, 0x12, 5, "d[4:0],rz[5:0],ry[5:0],by[3:0],bz[1],bx[4:0],gz[3:0],bz[0],gx[4:0],gy[3:0],rx[5:0],bz[4:3],bw[7:0],gy[4],bz[2],gw[7:0],by[4],gz[4],rw[7:0],m[4:0]",
8,5,5,5, 8,6,6,6, 8,5,5,5, 1, 0x16, 5, "d[4:0],bz[3],rz[4:0],bz[2],ry[4:0],by[3:0],bz[1],bx[4:0],gz[3:0],gx[5:0],gy[3:0],gz[4],rx[4:0],bz[4],gz[5],bw[7:0],gy[4],gy[5],gw[7:0],by[4],bz[0],rw[7:0],m[4:0]",
8,5,5,5, 8,5,5,5, 8,6,6,6, 1, 0x1a, 5, "d[4:0],bz[3],rz[4:0],bz[2],ry[4:0],by[3:0],bx[5:0],gz[3:0],bz[0],gx[4:0],gy[3:0],gz[4],rx[4:0],bz[4],bz[5],bw[7:0],gy[4],by[5],gw[7:0],by[4],bz[1],rw[7:0],m[4:0]",
7,6,6,6, 7,6,6,6, 7,6,6,6, 1, 0x01, 2, "d[4:0],rz[5:0],ry[5:0],by[3:0],bx[5:0],gz[3:0],gx[5:0],gy[3:0],rx[5:0],bz[4],bz[5],bz[3],bw[6:0],gy[4],bz[2],by[5],gw[6:0],by[4],bz[1:0],rw[6:0],gz[5:4],gy[5],m[1:0]",
6,6,6,6, 6,6,6,6, 6,6,6,6, 0, 0x1e, 5, "d[4:0],rz[5:0],ry[5:0],by[3:0],bx[5:0],gz[3:0],gx[5:0],gy[3:0],rx[5:0],bz[4],bz[5],bz[3],gz[5],bw[5:0],gy[4],bz[2],by[5],gy[5],gw[5:0],by[4],bz[1:0],gz[4],rw[5:0],m[4:0]",
};
// mapping of mode to the corresponding index in pattern
// UNUSED ZOH MODES are 0x13, 0x17, 0x1b, 0x1f -- return -2 for these
static const int mode_to_pat[MAXMODES] = {
3, // 0x00
8, // 0x01
0, // 0x02
-1,-1,-1,
1, // 0x06
-1,-1,-1,
2, // 0x0a
-1,-1,-1,
4, // 0x0e
-1,-1,-1,
5, // 0x12
-2,-1,-1,
6, // 0x16
-2,-1,-1,
7, // 0x1a
-2,-1,-1,
9, // 0x1e
-2
};
#define R_0(ep) (ep)[0].A[i]
#define R_1(ep) (ep)[0].B[i]
#define R_2(ep) (ep)[1].A[i]
#define R_3(ep) (ep)[1].B[i]
#define MASK(n) ((1<<(n))-1)
// compress endpoints
static void compress_endpts(const IntEndpts in[NREGIONS_TWO], ComprEndpts out[NREGIONS_TWO], const Pattern &p)
{
if (p.transformed)
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = R_0(in) & MASK(p.chan[i].prec[0]);
R_1(out) = (R_1(in) - R_0(in)) & MASK(p.chan[i].prec[1]);
R_2(out) = (R_2(in) - R_0(in)) & MASK(p.chan[i].prec[2]);
R_3(out) = (R_3(in) - R_0(in)) & MASK(p.chan[i].prec[3]);
}
}
else
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = R_0(in) & MASK(p.chan[i].prec[0]);
R_1(out) = R_1(in) & MASK(p.chan[i].prec[1]);
R_2(out) = R_2(in) & MASK(p.chan[i].prec[2]);
R_3(out) = R_3(in) & MASK(p.chan[i].prec[3]);
}
}
}
// decompress endpoints
static void decompress_endpts(const ComprEndpts in[NREGIONS_TWO], IntEndpts out[NREGIONS_TWO], const Pattern &p)
{
bool issigned = Utils::FORMAT == SIGNED_F16;
if (p.transformed)
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = issigned ? SIGN_EXTEND(R_0(in),p.chan[i].prec[0]) : R_0(in);
int t;
t = SIGN_EXTEND(R_1(in), p.chan[i].prec[1]);
t = (t + R_0(in)) & MASK(p.chan[i].prec[0]);
R_1(out) = issigned ? SIGN_EXTEND(t,p.chan[i].prec[0]) : t;
t = SIGN_EXTEND(R_2(in), p.chan[i].prec[2]);
t = (t + R_0(in)) & MASK(p.chan[i].prec[0]);
R_2(out) = issigned ? SIGN_EXTEND(t,p.chan[i].prec[0]) : t;
t = SIGN_EXTEND(R_3(in), p.chan[i].prec[3]);
t = (t + R_0(in)) & MASK(p.chan[i].prec[0]);
R_3(out) = issigned ? SIGN_EXTEND(t,p.chan[i].prec[0]) : t;
}
}
else
{
for (int i=0; i<NCHANNELS; ++i)
{
R_0(out) = issigned ? SIGN_EXTEND(R_0(in),p.chan[i].prec[0]) : R_0(in);
R_1(out) = issigned ? SIGN_EXTEND(R_1(in),p.chan[i].prec[1]) : R_1(in);
R_2(out) = issigned ? SIGN_EXTEND(R_2(in),p.chan[i].prec[2]) : R_2(in);
R_3(out) = issigned ? SIGN_EXTEND(R_3(in),p.chan[i].prec[3]) : R_3(in);
}
}
}
static void quantize_endpts(const FltEndpts endpts[NREGIONS_TWO], int prec, IntEndpts q_endpts[NREGIONS_TWO])
{
for (int region = 0; region < NREGIONS_TWO; ++region)
{
q_endpts[region].A[0] = Utils::quantize(endpts[region].A.x, prec);
q_endpts[region].A[1] = Utils::quantize(endpts[region].A.y, prec);
q_endpts[region].A[2] = Utils::quantize(endpts[region].A.z, prec);
q_endpts[region].B[0] = Utils::quantize(endpts[region].B.x, prec);
q_endpts[region].B[1] = Utils::quantize(endpts[region].B.y, prec);
q_endpts[region].B[2] = Utils::quantize(endpts[region].B.z, prec);
}
}
// swap endpoints as needed to ensure that the indices at index_positions have a 0 high-order bit
static void swap_indices(IntEndpts endpts[NREGIONS_TWO], int indices[Tile::TILE_H][Tile::TILE_W], int shapeindex)
{
for (int region = 0; region < NREGIONS_TWO; ++region)
{
int position = SHAPEINDEX_TO_COMPRESSED_INDICES(shapeindex,region);
int x = POS_TO_X(position);
int y = POS_TO_Y(position);
nvDebugCheck(REGION(x,y,shapeindex) == region); // double check the table
if (indices[y][x] & HIGH_INDEXBIT)
{
// high bit is set, swap the endpts and indices for this region
int t;
for (int i=0; i<NCHANNELS; ++i)
{
t = endpts[region].A[i]; endpts[region].A[i] = endpts[region].B[i]; endpts[region].B[i] = t;
}
for (int y = 0; y < Tile::TILE_H; y++)
for (int x = 0; x < Tile::TILE_W; x++)
if (REGION(x,y,shapeindex) == region)
indices[y][x] = NINDICES - 1 - indices[y][x];
}
}
}
// endpoints fit only if the compression was lossless
static bool endpts_fit(const IntEndpts orig[NREGIONS_TWO], const ComprEndpts compressed[NREGIONS_TWO], const Pattern &p)
{
IntEndpts uncompressed[NREGIONS_TWO];
decompress_endpts(compressed, uncompressed, p);
for (int j=0; j<NREGIONS_TWO; ++j)
{
for (int i=0; i<NCHANNELS; ++i)
{
if (orig[j].A[i] != uncompressed[j].A[i]) return false;
if (orig[j].B[i] != uncompressed[j].B[i]) return false;
}
}
return true;
}
static void write_header(const ComprEndpts endpts[NREGIONS_TWO], int shapeindex, const Pattern &p, Bits &out)
{
// interpret the verilog backwards and process it
int m = p.mode;
int d = shapeindex;
int rw = endpts[0].A[0], rx = endpts[0].B[0], ry = endpts[1].A[0], rz = endpts[1].B[0];
int gw = endpts[0].A[1], gx = endpts[0].B[1], gy = endpts[1].A[1], gz = endpts[1].B[1];
int bw = endpts[0].A[2], bx = endpts[0].B[2], by = endpts[1].A[2], bz = endpts[1].B[2];
int ptr = int(strlen(p.encoding));
while (ptr)
{
Field field;
int endbit, len;
// !!!UNDONE: get rid of string parsing!!!
Utils::parse(p.encoding, ptr, field, endbit, len);
switch(field)
{
case FIELD_M: out.write( m >> endbit, len); break;
case FIELD_D: out.write( d >> endbit, len); break;
case FIELD_RW: out.write(rw >> endbit, len); break;
case FIELD_RX: out.write(rx >> endbit, len); break;
case FIELD_RY: out.write(ry >> endbit, len); break;
case FIELD_RZ: out.write(rz >> endbit, len); break;
case FIELD_GW: out.write(gw >> endbit, len); break;
case FIELD_GX: out.write(gx >> endbit, len); break;
case FIELD_GY: out.write(gy >> endbit, len); break;
case FIELD_GZ: out.write(gz >> endbit, len); break;
case FIELD_BW: out.write(bw >> endbit, len); break;
case FIELD_BX: out.write(bx >> endbit, len); break;
case FIELD_BY: out.write(by >> endbit, len); break;
case FIELD_BZ: out.write(bz >> endbit, len); break;
default: nvUnreachable();
}
}
}
static bool read_header(Bits &in, ComprEndpts endpts[NREGIONS_TWO], int &shapeindex, Pattern &p)
{
// reading isn't quite symmetric with writing -- we don't know the encoding until we decode the mode
int mode = in.read(2);
if (mode != 0x00 && mode != 0x01)
mode = (in.read(3) << 2) | mode;
int pat_index = mode_to_pat[mode];
if (pat_index == -2)
return false; // reserved mode found
nvDebugCheck (pat_index >= 0 && pat_index < NPATTERNS);
nvDebugCheck (in.getptr() == patterns[pat_index].modebits);
p = patterns[pat_index];
int d;
int rw, rx, ry, rz;
int gw, gx, gy, gz;
int bw, bx, by, bz;
d = 0;
rw = rx = ry = rz = 0;
gw = gx = gy = gz = 0;
bw = bx = by = bz = 0;
int ptr = int(strlen(p.encoding));
while (ptr)
{
Field field;
int endbit, len;
// !!!UNDONE: get rid of string parsing!!!
Utils::parse(p.encoding, ptr, field, endbit, len);
switch(field)
{
case FIELD_M: break; // already processed so ignore
case FIELD_D: d |= in.read(len) << endbit; break;
case FIELD_RW: rw |= in.read(len) << endbit; break;
case FIELD_RX: rx |= in.read(len) << endbit; break;
case FIELD_RY: ry |= in.read(len) << endbit; break;
case FIELD_RZ: rz |= in.read(len) << endbit; break;
case FIELD_GW: gw |= in.read(len) << endbit; break;
case FIELD_GX: gx |= in.read(len) << endbit; break;
case FIELD_GY: gy |= in.read(len) << endbit; break;
case FIELD_GZ: gz |= in.read(len) << endbit; break;
case FIELD_BW: bw |= in.read(len) << endbit; break;
case FIELD_BX: bx |= in.read(len) << endbit; break;
case FIELD_BY: by |= in.read(len) << endbit; break;
case FIELD_BZ: bz |= in.read(len) << endbit; break;
default: nvUnreachable();
}
}
nvDebugCheck (in.getptr() == 128 - 46);
shapeindex = d;
endpts[0].A[0] = rw; endpts[0].B[0] = rx; endpts[1].A[0] = ry; endpts[1].B[0] = rz;
endpts[0].A[1] = gw; endpts[0].B[1] = gx; endpts[1].A[1] = gy; endpts[1].B[1] = gz;
endpts[0].A[2] = bw; endpts[0].B[2] = bx; endpts[1].A[2] = by; endpts[1].B[2] = bz;
return true;
}
static void write_indices(const int indices[Tile::TILE_H][Tile::TILE_W], int shapeindex, Bits &out)
{
int positions[NREGIONS_TWO];
for (int r = 0; r < NREGIONS_TWO; ++r)
positions[r] = SHAPEINDEX_TO_COMPRESSED_INDICES(shapeindex,r);
for (int pos = 0; pos < Tile::TILE_TOTAL; ++pos)
{
int x = POS_TO_X(pos);
int y = POS_TO_Y(pos);
bool match = false;
for (int r = 0; r < NREGIONS_TWO; ++r)
if (positions[r] == pos) { match = true; break; }
out.write(indices[y][x], INDEXBITS - (match ? 1 : 0));
}
}
static void emit_block(const ComprEndpts compr_endpts[NREGIONS_TWO], int shapeindex, const Pattern &p, const int indices[Tile::TILE_H][Tile::TILE_W], char *block)
{
Bits out(block, ZOH::BITSIZE);
write_header(compr_endpts, shapeindex, p, out);
write_indices(indices, shapeindex, out);
nvDebugCheck(out.getptr() == ZOH::BITSIZE);
}
static void generate_palette_quantized(const IntEndpts &endpts, int prec, Vector3 palette[NINDICES])
{
// scale endpoints
int a, b; // really need a IntVector3...
a = Utils::unquantize(endpts.A[0], prec);
b = Utils::unquantize(endpts.B[0], prec);
// interpolate
for (int i = 0; i < NINDICES; ++i)
palette[i].x = float(Utils::finish_unquantize(Utils::lerp(a, b, i, DENOM), prec));
a = Utils::unquantize(endpts.A[1], prec);
b = Utils::unquantize(endpts.B[1], prec);
// interpolate
for (int i = 0; i < NINDICES; ++i)
palette[i].y = float(Utils::finish_unquantize(Utils::lerp(a, b, i, DENOM), prec));
a = Utils::unquantize(endpts.A[2], prec);
b = Utils::unquantize(endpts.B[2], prec);
// interpolate
for (int i = 0; i < NINDICES; ++i)
palette[i].z = float(Utils::finish_unquantize(Utils::lerp(a, b, i, DENOM), prec));
}
static void read_indices(Bits &in, int shapeindex, int indices[Tile::TILE_H][Tile::TILE_W])
{
int positions[NREGIONS_TWO];
for (int r = 0; r < NREGIONS_TWO; ++r)
positions[r] = SHAPEINDEX_TO_COMPRESSED_INDICES(shapeindex,r);
for (int pos = 0; pos < Tile::TILE_TOTAL; ++pos)
{
int x = POS_TO_X(pos);
int y = POS_TO_Y(pos);
bool match = false;
for (int r = 0; r < NREGIONS_TWO; ++r)
if (positions[r] == pos) { match = true; break; }
indices[y][x]= in.read(INDEXBITS - (match ? 1 : 0));
}
}
void ZOH::decompresstwo(const char *block, Tile &t)
{
Bits in(block, ZOH::BITSIZE);
Pattern p;
IntEndpts endpts[NREGIONS_TWO];
ComprEndpts compr_endpts[NREGIONS_TWO];
int shapeindex;
if (!read_header(in, compr_endpts, shapeindex, p))
{
// reserved mode, return all zeroes
for (int y = 0; y < Tile::TILE_H; y++)
for (int x = 0; x < Tile::TILE_W; x++)
t.data[y][x] = Vector3(0.0f);
return;
}
decompress_endpts(compr_endpts, endpts, p);
Vector3 palette[NREGIONS_TWO][NINDICES];
for (int r = 0; r < NREGIONS_TWO; ++r)
generate_palette_quantized(endpts[r], p.chan[0].prec[0], &palette[r][0]);
int indices[Tile::TILE_H][Tile::TILE_W];
read_indices(in, shapeindex, indices);
nvDebugCheck(in.getptr() == ZOH::BITSIZE);
// lookup
for (int y = 0; y < Tile::TILE_H; y++)
for (int x = 0; x < Tile::TILE_W; x++)
t.data[y][x] = palette[REGION(x,y,shapeindex)][indices[y][x]];
}
// given a collection of colors and quantized endpoints, generate a palette, choose best entries, and return a single toterr
static float map_colors(const Vector3 colors[], const float importance[], int np, const IntEndpts &endpts, int prec)
{
Vector3 palette[NINDICES];
float toterr = 0;
Vector3 err;
generate_palette_quantized(endpts, prec, palette);
for (int i = 0; i < np; ++i)
{
float err, besterr;
besterr = Utils::norm(colors[i], palette[0]) * importance[i];
for (int j = 1; j < NINDICES && besterr > 0; ++j)
{
err = Utils::norm(colors[i], palette[j]) * importance[i];
if (err > besterr) // error increased, so we're done searching
break;
if (err < besterr)
besterr = err;
}
toterr += besterr;
}
return toterr;
}
// assign indices given a tile, shape, and quantized endpoints, return toterr for each region
static void assign_indices(const Tile &tile, int shapeindex, IntEndpts endpts[NREGIONS_TWO], int prec,
int indices[Tile::TILE_H][Tile::TILE_W], float toterr[NREGIONS_TWO])
{
// build list of possibles
Vector3 palette[NREGIONS_TWO][NINDICES];
for (int region = 0; region < NREGIONS_TWO; ++region)
{
generate_palette_quantized(endpts[region], prec, &palette[region][0]);
toterr[region] = 0;
}
Vector3 err;
for (int y = 0; y < tile.size_y; y++)
for (int x = 0; x < tile.size_x; x++)
{
int region = REGION(x,y,shapeindex);
float err, besterr;
besterr = Utils::norm(tile.data[y][x], palette[region][0]);
indices[y][x] = 0;
for (int i = 1; i < NINDICES && besterr > 0; ++i)
{
err = Utils::norm(tile.data[y][x], palette[region][i]);
if (err > besterr) // error increased, so we're done searching
break;
if (err < besterr)
{
besterr = err;
indices[y][x] = i;
}
}
toterr[region] += besterr;
}
}
static float perturb_one(const Vector3 colors[], const float importance[], int np, int ch, int prec, const IntEndpts &old_endpts, IntEndpts &new_endpts,
float old_err, int do_b)
{
// we have the old endpoints: old_endpts
// we have the perturbed endpoints: new_endpts
// we have the temporary endpoints: temp_endpts
IntEndpts temp_endpts;
float min_err = old_err; // start with the best current error
int beststep;
// copy real endpoints so we can perturb them
for (int i=0; i<NCHANNELS; ++i) { temp_endpts.A[i] = new_endpts.A[i] = old_endpts.A[i]; temp_endpts.B[i] = new_endpts.B[i] = old_endpts.B[i]; }
// do a logarithmic search for the best error for this endpoint (which)
for (int step = 1 << (prec-1); step; step >>= 1)
{
bool improved = false;
for (int sign = -1; sign <= 1; sign += 2)
{
if (do_b == 0)
{
temp_endpts.A[ch] = new_endpts.A[ch] + sign * step;
if (temp_endpts.A[ch] < 0 || temp_endpts.A[ch] >= (1 << prec))
continue;
}
else
{
temp_endpts.B[ch] = new_endpts.B[ch] + sign * step;
if (temp_endpts.B[ch] < 0 || temp_endpts.B[ch] >= (1 << prec))
continue;
}
float err = map_colors(colors, importance, np, temp_endpts, prec);
if (err < min_err)
{
improved = true;
min_err = err;
beststep = sign * step;
}
}
// if this was an improvement, move the endpoint and continue search from there
if (improved)
{
if (do_b == 0)
new_endpts.A[ch] += beststep;
else
new_endpts.B[ch] += beststep;
}
}
return min_err;
}
static void optimize_one(const Vector3 colors[], const float importance[], int np, float orig_err, const IntEndpts &orig_endpts, int prec, IntEndpts &opt_endpts)
{
float opt_err = orig_err;
for (int ch = 0; ch < NCHANNELS; ++ch)
{
opt_endpts.A[ch] = orig_endpts.A[ch];
opt_endpts.B[ch] = orig_endpts.B[ch];
}
/*
err0 = perturb(rgb0, delta0)
err1 = perturb(rgb1, delta1)
if (err0 < err1)
if (err0 >= initial_error) break
rgb0 += delta0
next = 1
else
if (err1 >= initial_error) break
rgb1 += delta1
next = 0
initial_err = map()
for (;;)
err = perturb(next ? rgb1:rgb0, delta)
if (err >= initial_err) break
next? rgb1 : rgb0 += delta
initial_err = err
*/
IntEndpts new_a, new_b;
IntEndpts new_endpt;
int do_b;
// now optimize each channel separately
for (int ch = 0; ch < NCHANNELS; ++ch)
{
// figure out which endpoint when perturbed gives the most improvement and start there
// if we just alternate, we can easily end up in a local minima
float err0 = perturb_one(colors, importance, np, ch, prec, opt_endpts, new_a, opt_err, 0); // perturb endpt A
float err1 = perturb_one(colors, importance, np, ch, prec, opt_endpts, new_b, opt_err, 1); // perturb endpt B
if (err0 < err1)
{
if (err0 >= opt_err)
continue;
opt_endpts.A[ch] = new_a.A[ch];
opt_err = err0;
do_b = 1; // do B next
}
else
{
if (err1 >= opt_err)
continue;
opt_endpts.B[ch] = new_b.B[ch];
opt_err = err1;
do_b = 0; // do A next
}
// now alternate endpoints and keep trying until there is no improvement
for (;;)
{
float err = perturb_one(colors, importance, np, ch, prec, opt_endpts, new_endpt, opt_err, do_b);
if (err >= opt_err)
break;
if (do_b == 0)
opt_endpts.A[ch] = new_endpt.A[ch];
else
opt_endpts.B[ch] = new_endpt.B[ch];
opt_err = err;
do_b = 1 - do_b; // now move the other endpoint
}
}
}
static void optimize_endpts(const Tile &tile, int shapeindex, const float orig_err[NREGIONS_TWO],
const IntEndpts orig_endpts[NREGIONS_TWO], int prec, IntEndpts opt_endpts[NREGIONS_TWO])
{
Vector3 pixels[Tile::TILE_TOTAL];
float importance[Tile::TILE_TOTAL];
float err = 0;
for (int region=0; region<NREGIONS_TWO; ++region)
{
// collect the pixels in the region
int np = 0;
for (int y = 0; y < tile.size_y; y++)
for (int x = 0; x < tile.size_x; x++)
if (REGION(x,y,shapeindex) == region)
{
pixels[np] = tile.data[y][x];
importance[np] = tile.importance_map[y][x];
++np;
}
optimize_one(pixels, importance, np, orig_err[region], orig_endpts[region], prec, opt_endpts[region]);
}
}
/* optimization algorithm
for each pattern
convert endpoints using pattern precision
assign indices and get initial error
compress indices (and possibly reorder endpoints)
transform endpoints
if transformed endpoints fit pattern
get original endpoints back
optimize endpoints, get new endpoints, new indices, and new error // new error will almost always be better
compress new indices
transform new endpoints
if new endpoints fit pattern AND if error is improved
emit compressed block with new data
else
emit compressed block with original data // to try to preserve maximum endpoint precision
*/
float ZOH::refinetwo(const Tile &tile, int shapeindex_best, const FltEndpts endpts[NREGIONS_TWO], char *block)
{
float orig_err[NREGIONS_TWO], opt_err[NREGIONS_TWO], orig_toterr, opt_toterr;
IntEndpts orig_endpts[NREGIONS_TWO], opt_endpts[NREGIONS_TWO];
ComprEndpts compr_orig[NREGIONS_TWO], compr_opt[NREGIONS_TWO];
int orig_indices[Tile::TILE_H][Tile::TILE_W], opt_indices[Tile::TILE_H][Tile::TILE_W];
for (int sp = 0; sp < NPATTERNS; ++sp)
{
// precisions for all channels need to be the same
for (int i=1; i<NCHANNELS; ++i) nvDebugCheck (patterns[sp].chan[0].prec[0] == patterns[sp].chan[i].prec[0]);
quantize_endpts(endpts, patterns[sp].chan[0].prec[0], orig_endpts);
assign_indices(tile, shapeindex_best, orig_endpts, patterns[sp].chan[0].prec[0], orig_indices, orig_err);
swap_indices(orig_endpts, orig_indices, shapeindex_best);
compress_endpts(orig_endpts, compr_orig, patterns[sp]);
if (endpts_fit(orig_endpts, compr_orig, patterns[sp]))
{
optimize_endpts(tile, shapeindex_best, orig_err, orig_endpts, patterns[sp].chan[0].prec[0], opt_endpts);
assign_indices(tile, shapeindex_best, opt_endpts, patterns[sp].chan[0].prec[0], opt_indices, opt_err);
swap_indices(opt_endpts, opt_indices, shapeindex_best);
compress_endpts(opt_endpts, compr_opt, patterns[sp]);
orig_toterr = opt_toterr = 0;
for (int i=0; i < NREGIONS_TWO; ++i) { orig_toterr += orig_err[i]; opt_toterr += opt_err[i]; }
if (endpts_fit(opt_endpts, compr_opt, patterns[sp]) && opt_toterr < orig_toterr)
{
emit_block(compr_opt, shapeindex_best, patterns[sp], opt_indices, block);
return opt_toterr;
}
else
{
// either it stopped fitting when we optimized it, or there was no improvement
// so go back to the unoptimized endpoints which we know will fit
emit_block(compr_orig, shapeindex_best, patterns[sp], orig_indices, block);
return orig_toterr;
}
}
}
nvAssert(false); //throw "No candidate found, should never happen (refinetwo.)";
return FLT_MAX;
}
static void generate_palette_unquantized(const FltEndpts endpts[NREGIONS_TWO], Vector3 palette[NREGIONS_TWO][NINDICES])
{
for (int region = 0; region < NREGIONS_TWO; ++region)
for (int i = 0; i < NINDICES; ++i)
palette[region][i] = Utils::lerp(endpts[region].A, endpts[region].B, i, DENOM);
}
// generate a palette from unquantized endpoints, then pick best palette color for all pixels in each region, return toterr for all regions combined
static float map_colors(const Tile &tile, int shapeindex, const FltEndpts endpts[NREGIONS_TWO])
{
// build list of possibles
Vector3 palette[NREGIONS_TWO][NINDICES];
generate_palette_unquantized(endpts, palette);
float toterr = 0;
Vector3 err;
for (int y = 0; y < tile.size_y; y++)
for (int x = 0; x < tile.size_x; x++)
{
int region = REGION(x,y,shapeindex);
float err, besterr;
besterr = Utils::norm(tile.data[y][x], palette[region][0]) * tile.importance_map[y][x];
for (int i = 1; i < NINDICES && besterr > 0; ++i)
{
err = Utils::norm(tile.data[y][x], palette[region][i]) * tile.importance_map[y][x];
if (err > besterr) // error increased, so we're done searching
break;
if (err < besterr)
besterr = err;
}
toterr += besterr;
}
return toterr;
}
float ZOH::roughtwo(const Tile &tile, int shapeindex, FltEndpts endpts[NREGIONS_TWO])
{
for (int region=0; region<NREGIONS_TWO; ++region)
{
int np = 0;
Vector3 colors[Tile::TILE_TOTAL];
Vector3 mean(0,0,0);
for (int y = 0; y < tile.size_y; y++)
for (int x = 0; x < tile.size_x; x++)
if (REGION(x,y,shapeindex) == region)
{
colors[np] = tile.data[y][x];
mean += tile.data[y][x];
++np;
}
// handle simple cases
if (np == 0)
{
Vector3 zero(0,0,0);
endpts[region].A = zero;
endpts[region].B = zero;
continue;
}
else if (np == 1)
{
endpts[region].A = colors[0];
endpts[region].B = colors[0];
continue;
}
else if (np == 2)
{
endpts[region].A = colors[0];
endpts[region].B = colors[1];
continue;
}
mean /= float(np);
Vector3 direction = Fit::computePrincipalComponent_EigenSolver(np, colors);
// project each pixel value along the principal direction
float minp = FLT_MAX, maxp = -FLT_MAX;
for (int i = 0; i < np; i++)
{
float dp = dot(colors[i]-mean, direction);
if (dp < minp) minp = dp;
if (dp > maxp) maxp = dp;
}
// choose as endpoints 2 points along the principal direction that span the projections of all of the pixel values
endpts[region].A = mean + minp*direction;
endpts[region].B = mean + maxp*direction;
// clamp endpoints
// the argument for clamping is that the actual endpoints need to be clamped and thus we need to choose the best
// shape based on endpoints being clamped
Utils::clamp(endpts[region].A);
Utils::clamp(endpts[region].B);
}
return map_colors(tile, shapeindex, endpts);
}
float ZOH::compresstwo(const Tile &t, char *block)
{
int shapeindex_best = 0;
FltEndpts endptsbest[NREGIONS_TWO], tempendpts[NREGIONS_TWO];
float msebest = FLT_MAX;
/*
collect the mse values that are within 5% of the best values
optimize each one and choose the best
*/
// hack for now -- just use the best value WORK
for (int i=0; i<NSHAPES && msebest>0.0; ++i)
{
float mse = roughtwo(t, i, tempendpts);
if (mse < msebest)
{
msebest = mse;
shapeindex_best = i;
memcpy(endptsbest, tempendpts, sizeof(endptsbest));
}
}
return refinetwo(t, shapeindex_best, endptsbest, block);
}