2006-04-24 01:14:18 +02:00
|
|
|
/**
|
|
|
|
* =========================================================================
|
|
|
|
* File : tex_dds.cpp
|
|
|
|
* Project : 0 A.D.
|
|
|
|
* Description : DDS (DirectDraw Surface) codec.
|
|
|
|
* =========================================================================
|
|
|
|
*/
|
|
|
|
|
2007-05-07 18:33:24 +02:00
|
|
|
// license: GPL; see lib/license.txt
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
#include "precompiled.h"
|
|
|
|
|
|
|
|
#include "lib/byte_order.h"
|
|
|
|
#include "tex_codec.h"
|
2007-05-09 23:01:11 +02:00
|
|
|
#include "lib/bits.h"
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// NOTE: the convention is bottom-up for DDS, but there's no way to tell.
|
|
|
|
|
|
|
|
//-----------------------------------------------------------------------------
|
|
|
|
// S3TC decompression
|
|
|
|
//-----------------------------------------------------------------------------
|
|
|
|
|
|
|
|
// note: this code is not so efficient (mostly due to splitting it up
|
|
|
|
// into function calls for readability). that's because it's only used to
|
|
|
|
// emulate hardware S3TC support - if that isn't available, everything will
|
|
|
|
// be dog-slow anyway due to increased vmem usage.
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
// pixel colors are stored as size_t[4]. size_t rather than u8 protects from
|
2006-04-24 01:14:18 +02:00
|
|
|
// overflow during calculations, and padding to an even size is a bit
|
|
|
|
// more efficient (even though we don't need the alpha component).
|
|
|
|
enum RGBA { R, G, B, A };
|
|
|
|
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static inline void mix_2_3(size_t dst[4], size_t c0[4], size_t c1[4])
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
for(int i = 0; i < 3; i++) dst[i] = (c0[i]*2 + c1[i] + 1)/3;
|
|
|
|
}
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static inline void mix_avg(size_t dst[4], size_t c0[4], size_t c1[4])
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
for(int i = 0; i < 3; i++) dst[i] = (c0[i]+c1[i])/2;
|
|
|
|
}
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static inline size_t access_bit_tbl(u32 tbl, size_t idx, size_t bit_width)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t val = tbl >> (idx*bit_width);
|
|
|
|
val &= bit_mask<u32>(bit_width);
|
2006-04-24 01:14:18 +02:00
|
|
|
return val;
|
|
|
|
}
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static inline size_t access_bit_tbl64(u64 tbl, size_t idx, size_t bit_width)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t val = (size_t)(tbl >> (idx*bit_width));
|
|
|
|
val &= bit_mask<u64>(bit_width);
|
2006-04-24 01:14:18 +02:00
|
|
|
return val;
|
|
|
|
}
|
|
|
|
|
|
|
|
// extract a range of bits and expand to 8 bits (by replicating
|
|
|
|
// MS bits - see http://www.mindcontrol.org/~hplus/graphics/expand-bits.html ;
|
|
|
|
// this is also the algorithm used by graphics cards when decompressing S3TC).
|
|
|
|
// used to convert 565 to 32bpp RGB.
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static inline size_t unpack_to_8(u16 c, size_t bits_below, size_t num_bits)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t num_filler_bits = 8-num_bits;
|
|
|
|
const size_t field = (size_t)bits(c, bits_below, bits_below+num_bits-1);
|
|
|
|
const size_t filler = field >> (8-num_bits);
|
2006-04-24 01:14:18 +02:00
|
|
|
return (field << num_filler_bits) | filler;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
// for efficiency, we precalculate as much as possible about a block
|
|
|
|
// and store it here.
|
|
|
|
struct S3tcBlock
|
|
|
|
{
|
|
|
|
// the 4 color choices for each pixel (RGBA)
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t c[4][4]; // c[i][RGBA_component]
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// (DXT5 only) the 8 alpha choices
|
|
|
|
u8 dxt5_a_tbl[8];
|
|
|
|
|
|
|
|
// alpha block; interpretation depends on dxt.
|
|
|
|
u64 a_bits;
|
|
|
|
|
|
|
|
// table of 2-bit color selectors
|
|
|
|
u32 c_selectors;
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t dxt;
|
2006-04-24 01:14:18 +02:00
|
|
|
};
|
|
|
|
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static void s3tc_precalc_alpha(size_t dxt, const u8* RESTRICT a_block, S3tcBlock* RESTRICT b)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
// read block contents
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const u8 a0 = a_block[0], a1 = a_block[1];
|
2006-04-24 01:14:18 +02:00
|
|
|
b->a_bits = read_le64(a_block); // see below
|
|
|
|
|
|
|
|
if(dxt == 5)
|
|
|
|
{
|
|
|
|
// skip a0,a1 bytes (data is little endian)
|
|
|
|
b->a_bits >>= 16;
|
|
|
|
|
|
|
|
const bool is_dxt5_special_combination = (a0 <= a1);
|
|
|
|
u8* a = b->dxt5_a_tbl; // shorthand
|
|
|
|
if(is_dxt5_special_combination)
|
|
|
|
{
|
|
|
|
a[0] = a0;
|
|
|
|
a[1] = a1;
|
|
|
|
a[2] = (4*a0 + 1*a1 + 2)/5;
|
|
|
|
a[3] = (3*a0 + 2*a1 + 2)/5;
|
|
|
|
a[4] = (2*a0 + 3*a1 + 2)/5;
|
|
|
|
a[5] = (1*a0 + 4*a1 + 2)/5;
|
|
|
|
a[6] = 0;
|
|
|
|
a[7] = 255;
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
a[0] = a0;
|
|
|
|
a[1] = a1;
|
|
|
|
a[2] = (6*a0 + 1*a1 + 3)/7;
|
|
|
|
a[3] = (5*a0 + 2*a1 + 3)/7;
|
|
|
|
a[4] = (4*a0 + 3*a1 + 3)/7;
|
|
|
|
a[5] = (3*a0 + 4*a1 + 3)/7;
|
|
|
|
a[6] = (2*a0 + 5*a1 + 3)/7;
|
|
|
|
a[7] = (1*a0 + 6*a1 + 3)/7;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static void s3tc_precalc_color(size_t dxt, const u8* RESTRICT c_block, S3tcBlock* RESTRICT b)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
// read block contents
|
|
|
|
// .. S3TC reference colors (565 format). the color table is generated
|
|
|
|
// from some combination of these, depending on their ordering.
|
|
|
|
u16 rc[2];
|
|
|
|
for(int i = 0; i < 2; i++)
|
|
|
|
rc[i] = read_le16(c_block + 2*i);
|
|
|
|
// .. table of 2-bit color selectors
|
|
|
|
b->c_selectors = read_le32(c_block+4);
|
|
|
|
|
|
|
|
const bool is_dxt1_special_combination =
|
|
|
|
(dxt == 1 || dxt == DXT1A) && rc[0] <= rc[1];
|
|
|
|
|
|
|
|
// c0 and c1 are the values of rc[], converted to 32bpp
|
|
|
|
for(int i = 0; i < 2; i++)
|
|
|
|
{
|
|
|
|
b->c[i][R] = unpack_to_8(rc[i], 11, 5);
|
|
|
|
b->c[i][G] = unpack_to_8(rc[i], 5, 6);
|
|
|
|
b->c[i][B] = unpack_to_8(rc[i], 0, 5);
|
|
|
|
}
|
|
|
|
|
|
|
|
// c2 and c3 are combinations of c0 and c1:
|
|
|
|
if(is_dxt1_special_combination)
|
|
|
|
{
|
|
|
|
mix_avg(b->c[2], b->c[0], b->c[1]); // c2 = (c0+c1)/2
|
|
|
|
for(int i = 0; i < 3; i++) b->c[3][i] = 0; // c3 = black
|
|
|
|
b->c[3][A] = (dxt == DXT1A)? 0 : 255; // (transparent iff DXT1a)
|
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
mix_2_3(b->c[2], b->c[0], b->c[1]); // c2 = 2/3*c0 + 1/3*c1
|
|
|
|
mix_2_3(b->c[3], b->c[1], b->c[0]); // c3 = 1/3*c0 + 2/3*c1
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static void s3tc_precalc_block(size_t dxt, const u8* RESTRICT block, S3tcBlock* RESTRICT b)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
b->dxt = dxt;
|
|
|
|
|
|
|
|
// (careful, 'dxt != 1' doesn't work - there's also DXT1a)
|
|
|
|
const u8* a_block = block;
|
|
|
|
const u8* c_block = (dxt == 3 || dxt == 5)? block+8 : block;
|
|
|
|
|
|
|
|
s3tc_precalc_alpha(dxt, a_block, b);
|
|
|
|
s3tc_precalc_color(dxt, c_block, b);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static void s3tc_write_pixel(const S3tcBlock* RESTRICT b, size_t pixel_idx, u8* RESTRICT out)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
debug_assert(pixel_idx < 16);
|
|
|
|
|
|
|
|
// pixel index -> color selector (2 bit) -> color
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t c_selector = access_bit_tbl(b->c_selectors, pixel_idx, 2);
|
|
|
|
const size_t* c = b->c[c_selector];
|
2006-04-24 01:14:18 +02:00
|
|
|
for(int i = 0; i < 3; i++)
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
out[i] = (u8)c[i];
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// if no alpha, done
|
|
|
|
if(b->dxt == 1)
|
|
|
|
return;
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t a;
|
2006-04-24 01:14:18 +02:00
|
|
|
if(b->dxt == 3)
|
|
|
|
{
|
|
|
|
// table of 4-bit alpha entries
|
|
|
|
a = access_bit_tbl64(b->a_bits, pixel_idx, 4);
|
|
|
|
a |= a << 4; // expand to 8 bits (replicate high into low!)
|
|
|
|
}
|
|
|
|
else if(b->dxt == 5)
|
|
|
|
{
|
|
|
|
// pixel index -> alpha selector (3 bit) -> alpha
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t a_selector = access_bit_tbl64(b->a_bits, pixel_idx, 3);
|
2006-04-24 01:14:18 +02:00
|
|
|
a = b->dxt5_a_tbl[a_selector];
|
|
|
|
}
|
|
|
|
// (dxt == DXT1A)
|
|
|
|
else
|
|
|
|
a = c[A];
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
out[A] = (u8)a;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
struct S3tcDecompressInfo
|
|
|
|
{
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t dxt;
|
|
|
|
size_t s3tc_block_size;
|
|
|
|
size_t out_Bpp;
|
2006-04-24 01:14:18 +02:00
|
|
|
u8* out;
|
|
|
|
};
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static void s3tc_decompress_level(size_t UNUSED(level), size_t level_w, size_t level_h,
|
2007-09-25 12:43:11 +02:00
|
|
|
const u8* RESTRICT level_data, size_t level_data_size, void* RESTRICT cbData)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
2007-09-25 12:43:11 +02:00
|
|
|
S3tcDecompressInfo* di = (S3tcDecompressInfo*)cbData;
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t dxt = di->dxt;
|
|
|
|
const size_t s3tc_block_size = di->s3tc_block_size;
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// note: 1x1 images are legitimate (e.g. in mipmaps). they report their
|
|
|
|
// width as such for glTexImage, but the S3TC data is padded to
|
|
|
|
// 4x4 pixel block boundaries.
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t blocks_w = round_up(level_w, 4u) / 4u;
|
|
|
|
const size_t blocks_h = round_up(level_h, 4u) / 4u;
|
2006-04-24 01:14:18 +02:00
|
|
|
const u8* s3tc_data = level_data;
|
|
|
|
debug_assert(level_data_size % s3tc_block_size == 0);
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
for(size_t block_y = 0; block_y < blocks_h; block_y++)
|
|
|
|
for(size_t block_x = 0; block_x < blocks_w; block_x++)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
S3tcBlock b;
|
|
|
|
s3tc_precalc_block(dxt, s3tc_data, &b);
|
|
|
|
s3tc_data += s3tc_block_size;
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t pixel_idx = 0;
|
2006-04-24 01:14:18 +02:00
|
|
|
for(int y = 0; y < 4; y++)
|
|
|
|
{
|
|
|
|
// this is ugly, but advancing after x, y and block_y loops
|
|
|
|
// is no better.
|
|
|
|
u8* out = (u8*)di->out + ((block_y*4+y)*blocks_w*4 + block_x*4) * di->out_Bpp;
|
|
|
|
for(int x = 0; x < 4; x++)
|
|
|
|
{
|
|
|
|
s3tc_write_pixel(&b, pixel_idx, out);
|
|
|
|
out += di->out_Bpp;
|
|
|
|
pixel_idx++;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
} // for block_x
|
|
|
|
|
|
|
|
debug_assert(s3tc_data == level_data + level_data_size);
|
|
|
|
di->out += blocks_w*blocks_h * 16 * di->out_Bpp;
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
// decompress the given image (which is known to be stored as DXTn)
|
|
|
|
// effectively in-place. updates Tex fields.
|
|
|
|
static LibError s3tc_decompress(Tex* t)
|
|
|
|
{
|
|
|
|
// alloc new image memory
|
|
|
|
// notes:
|
|
|
|
// - dxt == 1 is the only non-alpha case.
|
|
|
|
// - adding or stripping alpha channels during transform is not
|
|
|
|
// our job; we merely output the same pixel format as given
|
|
|
|
// (tex.cpp's plain transform could cover it, if ever needed).
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t dxt = t->flags & TEX_DXT;
|
|
|
|
const size_t out_bpp = (dxt != 1)? 32 : 24;
|
2006-04-24 01:14:18 +02:00
|
|
|
const size_t out_size = tex_img_size(t) * out_bpp / t->bpp;
|
2007-12-20 21:14:21 +01:00
|
|
|
shared_ptr<u8> decompressedData = io_Allocate(out_size);
|
2006-04-24 01:14:18 +02:00
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t s3tc_block_size = (dxt == 3 || dxt == 5)? 16 : 8;
|
2007-12-20 21:14:21 +01:00
|
|
|
S3tcDecompressInfo di = { dxt, s3tc_block_size, out_bpp/8, decompressedData.get() };
|
2006-04-24 01:14:18 +02:00
|
|
|
const u8* s3tc_data = tex_get_data(t);
|
|
|
|
const int levels_to_skip = (t->flags & TEX_MIPMAPS)? 0 : TEX_BASE_LEVEL_ONLY;
|
|
|
|
tex_util_foreach_mipmap(t->w, t->h, t->bpp, s3tc_data, levels_to_skip, 4, s3tc_decompress_level, &di);
|
2007-12-20 21:14:21 +01:00
|
|
|
t->data = decompressedData;
|
|
|
|
t->dataSize = out_size;
|
2006-04-24 01:14:18 +02:00
|
|
|
t->ofs = 0;
|
|
|
|
t->bpp = out_bpp;
|
|
|
|
t->flags &= ~TEX_DXT;
|
2006-09-22 15:19:40 +02:00
|
|
|
return INFO::OK;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
//-----------------------------------------------------------------------------
|
|
|
|
// DDS file format
|
|
|
|
//-----------------------------------------------------------------------------
|
|
|
|
|
|
|
|
// bit values and structure definitions taken from
|
|
|
|
// http://msdn.microsoft.com/archive/en-us/directx9_c/directx/graphics/reference/DDSFileReference/ddsfileformat.asp
|
|
|
|
|
|
|
|
#pragma pack(push, 1)
|
|
|
|
|
|
|
|
// DDPIXELFORMAT.dwFlags
|
|
|
|
// we've seen some DXT3 files that don't have this set (which is nonsense;
|
|
|
|
// any image lacking alpha should be stored as DXT1). it's authoritative
|
|
|
|
// if fourcc is DXT1 (there's no other way to tell DXT1 and DXT1a apart)
|
|
|
|
// and ignored otherwise.
|
|
|
|
#define DDPF_ALPHAPIXELS 0x00000001
|
|
|
|
#define DDPF_FOURCC 0x00000004
|
|
|
|
#define DDPF_RGB 0x00000040
|
|
|
|
|
|
|
|
typedef struct
|
|
|
|
{
|
|
|
|
u32 dwSize; // size of structure (32)
|
|
|
|
u32 dwFlags; // indicates which fields are valid
|
|
|
|
u32 dwFourCC; // (DDPF_FOURCC) FOURCC code, "DXTn"
|
|
|
|
u32 dwRGBBitCount; // (DDPF_RGB) bits per pixel
|
|
|
|
u32 dwRBitMask;
|
|
|
|
u32 dwGBitMask;
|
|
|
|
u32 dwBBitMask;
|
|
|
|
u32 dwRGBAlphaBitMask;
|
|
|
|
}
|
|
|
|
DDPIXELFORMAT;
|
|
|
|
|
|
|
|
// DDCAPS2.dwCaps1
|
|
|
|
#define DDSCAPS_COMPLEX 0x00000008
|
|
|
|
#define DDSCAPS_TEXTURE 0x00001000
|
|
|
|
#define DDSCAPS_MIPMAP 0x00400000
|
|
|
|
|
|
|
|
// DDCAPS2.dwCaps2
|
|
|
|
#define DDSCAPS2_CUBEMAP 0x00000200
|
|
|
|
#define DDSCAPS2_CUBEMAP_POSITIVEX 0x00000400
|
|
|
|
#define DDSCAPS2_CUBEMAP_NEGATIVEX 0x00000800
|
|
|
|
#define DDSCAPS2_CUBEMAP_POSITIVEY 0x00001000
|
|
|
|
#define DDSCAPS2_CUBEMAP_NEGATIVEY 0x00002000
|
|
|
|
#define DDSCAPS2_CUBEMAP_POSITIVEZ 0x00004000
|
|
|
|
#define DDSCAPS2_CUBEMAP_NEGATIVEZ 0x00008000
|
|
|
|
#define DDSCAPS2_VOLUME 0x00200000
|
|
|
|
|
|
|
|
typedef struct
|
|
|
|
{
|
|
|
|
u32 dwCaps1;
|
|
|
|
u32 dwCaps2;
|
|
|
|
u32 Reserved[2];
|
|
|
|
}
|
|
|
|
DDCAPS2;
|
|
|
|
|
|
|
|
// DDSURFACEDESC2.dwFlags
|
|
|
|
#define DDSD_CAPS 0x00000001
|
|
|
|
#define DDSD_HEIGHT 0x00000002
|
|
|
|
#define DDSD_WIDTH 0x00000004
|
|
|
|
#define DDSD_PITCH 0x00000008
|
|
|
|
#define DDSD_PIXELFORMAT 0x00001000
|
|
|
|
#define DDSD_MIPMAPCOUNT 0x00020000
|
|
|
|
#define DDSD_LINEARSIZE 0x00080000
|
|
|
|
#define DDSD_DEPTH 0x00800000
|
|
|
|
|
|
|
|
typedef struct
|
|
|
|
{
|
|
|
|
u32 dwSize; // size of structure (124)
|
|
|
|
u32 dwFlags; // indicates which fields are valid
|
|
|
|
u32 dwHeight; // (DDSD_HEIGHT) height of main image (pixels)
|
|
|
|
u32 dwWidth; // (DDSD_WIDTH ) width of main image (pixels)
|
|
|
|
u32 dwPitchOrLinearSize; // (DDSD_LINEARSIZE) total image size
|
|
|
|
// (DDSD_PITCH) bytes per row (%4 = 0)
|
|
|
|
u32 dwDepth; // (DDSD_DEPTH) vol. textures: vol. depth
|
|
|
|
u32 dwMipMapCount; // (DDSD_MIPMAPCOUNT) total # levels
|
|
|
|
u32 dwReserved1[11]; // reserved
|
|
|
|
DDPIXELFORMAT ddpfPixelFormat; // (DDSD_PIXELFORMAT) surface description
|
|
|
|
DDCAPS2 ddsCaps; // (DDSD_CAPS) misc. surface flags
|
|
|
|
u32 dwReserved2; // reserved
|
|
|
|
}
|
|
|
|
DDSURFACEDESC2;
|
|
|
|
|
|
|
|
#pragma pack(pop)
|
|
|
|
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static bool is_valid_dxt(size_t dxt)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
switch(dxt)
|
|
|
|
{
|
|
|
|
case 0:
|
|
|
|
case 1:
|
|
|
|
case DXT1A:
|
|
|
|
case 3:
|
|
|
|
case 5:
|
|
|
|
return true;
|
|
|
|
default:
|
|
|
|
return false;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
// extract all information from DDS pixel format and store in bpp, flags.
|
|
|
|
// pf points to the DDS file's header; all fields must be endian-converted
|
|
|
|
// before use.
|
|
|
|
// output parameters invalid on failure.
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static LibError decode_pf(const DDPIXELFORMAT* pf, size_t& bpp, size_t& flags)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
bpp = 0;
|
|
|
|
flags = 0;
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// check struct size
|
|
|
|
if(read_le32(&pf->dwSize) != sizeof(DDPIXELFORMAT))
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::TEX_INVALID_SIZE);
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// determine type
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t pf_flags = (size_t)read_le32(&pf->dwFlags);
|
2006-04-24 01:14:18 +02:00
|
|
|
// .. uncompressed
|
|
|
|
if(pf_flags & DDPF_RGB)
|
|
|
|
{
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t pf_bpp = (size_t)read_le32(&pf->dwRGBBitCount);
|
|
|
|
const size_t pf_r_mask = (size_t)read_le32(&pf->dwRBitMask);
|
|
|
|
const size_t pf_g_mask = (size_t)read_le32(&pf->dwGBitMask);
|
|
|
|
const size_t pf_b_mask = (size_t)read_le32(&pf->dwBBitMask);
|
|
|
|
const size_t pf_a_mask = (size_t)read_le32(&pf->dwRGBAlphaBitMask);
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// (checked below; must be set in case below warning is to be
|
|
|
|
// skipped)
|
|
|
|
bpp = pf_bpp;
|
|
|
|
|
|
|
|
if(pf_flags & DDPF_ALPHAPIXELS)
|
|
|
|
{
|
|
|
|
// something weird other than RGBA or BGRA
|
|
|
|
if(pf_a_mask != 0xFF000000)
|
|
|
|
goto unsupported_component_ordering;
|
|
|
|
flags |= TEX_ALPHA;
|
|
|
|
}
|
|
|
|
|
|
|
|
// make sure component ordering is 0xBBGGRR = RGB (see below)
|
|
|
|
if(pf_r_mask != 0xFF || pf_g_mask != 0xFF00 || pf_b_mask != 0xFF0000)
|
|
|
|
{
|
|
|
|
// DDPIXELFORMAT in theory supports any ordering of R,G,B,A.
|
|
|
|
// we need to upload to OpenGL, which can only receive BGR(A) or
|
|
|
|
// RGB(A). the former still requires conversion (done by driver),
|
|
|
|
// so it's slower. since the very purpose of supporting uncompressed
|
|
|
|
// DDS is storing images in a format that requires no processing,
|
|
|
|
// we do not allow any weird orderings that require runtime work.
|
|
|
|
// instead, the artists must export with the correct settings.
|
|
|
|
unsupported_component_ordering:
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::TEX_FMT_INVALID);
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
RETURN_ERR(tex_validate_plain_format(bpp, flags));
|
|
|
|
}
|
|
|
|
// .. compressed
|
|
|
|
else if(pf_flags & DDPF_FOURCC)
|
|
|
|
{
|
|
|
|
// set effective bpp and store DXT format in flags & TEX_DXT.
|
|
|
|
// no endian conversion necessary - FOURCC() takes care of that.
|
|
|
|
switch(pf->dwFourCC)
|
|
|
|
{
|
|
|
|
case FOURCC('D','X','T','1'):
|
|
|
|
bpp = 4;
|
|
|
|
if(pf_flags & DDPF_ALPHAPIXELS)
|
|
|
|
flags |= DXT1A | TEX_ALPHA;
|
|
|
|
else
|
|
|
|
flags |= 1;
|
|
|
|
break;
|
|
|
|
case FOURCC('D','X','T','3'):
|
|
|
|
bpp = 8;
|
|
|
|
flags |= 3;
|
|
|
|
flags |= TEX_ALPHA; // see DDPF_ALPHAPIXELS decl
|
|
|
|
break;
|
|
|
|
case FOURCC('D','X','T','5'):
|
|
|
|
bpp = 8;
|
|
|
|
flags |= 5;
|
|
|
|
flags |= TEX_ALPHA; // see DDPF_ALPHAPIXELS decl
|
|
|
|
break;
|
|
|
|
|
|
|
|
default:
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::TEX_FMT_INVALID);
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
}
|
|
|
|
// .. neither uncompressed nor compressed - invalid
|
|
|
|
else
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::TEX_FMT_INVALID);
|
2006-04-24 01:14:18 +02:00
|
|
|
|
2006-09-22 15:19:40 +02:00
|
|
|
return INFO::OK;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
// extract all information from DDS header and store in w, h, bpp, flags.
|
|
|
|
// sd points to the DDS file's header; all fields must be endian-converted
|
|
|
|
// before use.
|
|
|
|
// output parameters invalid on failure.
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static LibError decode_sd(const DDSURFACEDESC2* sd, size_t* w_, size_t* h_,
|
|
|
|
size_t* bpp_, size_t* flags_)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
// check header size
|
|
|
|
if(read_le32(&sd->dwSize) != sizeof(*sd))
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::CORRUPTED);
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// flags (indicate which fields are valid)
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t sd_flags = (size_t)read_le32(&sd->dwFlags);
|
2006-04-24 01:14:18 +02:00
|
|
|
// .. not all required fields are present
|
|
|
|
// note: we can't guess dimensions - the image may not be square.
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t sd_req_flags = DDSD_CAPS|DDSD_HEIGHT|DDSD_WIDTH|DDSD_PIXELFORMAT;
|
2006-04-24 01:14:18 +02:00
|
|
|
if((sd_flags & sd_req_flags) != sd_req_flags)
|
2007-12-20 21:14:21 +01:00
|
|
|
WARN_RETURN(ERR::TEX_INCOMPLETE_HEADER);
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// image dimensions
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t h = (size_t)read_le32(&sd->dwHeight);
|
|
|
|
const size_t w = (size_t)read_le32(&sd->dwWidth);
|
2006-04-24 01:14:18 +02:00
|
|
|
|
|
|
|
// pixel format
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t bpp, flags;
|
|
|
|
RETURN_ERR(decode_pf(&sd->ddpfPixelFormat, bpp, flags));
|
2006-04-24 01:14:18 +02:00
|
|
|
|
2006-12-15 17:09:30 +01:00
|
|
|
// if the image is not aligned with the S3TC block size, it is stored
|
|
|
|
// with extra pixels on the bottom left to fill up the space, so we need
|
|
|
|
// to account for those when calculating how big it should be
|
2008-05-01 17:41:42 +02:00
|
|
|
size_t stored_h, stored_w;
|
2006-12-15 17:09:30 +01:00
|
|
|
if(flags & TEX_DXT)
|
|
|
|
{
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
stored_h = round_up(h, size_t(4));
|
|
|
|
stored_w = round_up(w, size_t(4));
|
2006-12-15 17:09:30 +01:00
|
|
|
}
|
|
|
|
else
|
|
|
|
{
|
|
|
|
stored_h = h;
|
|
|
|
stored_w = w;
|
|
|
|
}
|
|
|
|
|
2006-04-24 01:14:18 +02:00
|
|
|
// verify pitch or linear size, if given
|
2006-12-15 17:09:30 +01:00
|
|
|
const size_t pitch = stored_w*bpp/8;
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t sd_pitch_or_size = (size_t)read_le32(&sd->dwPitchOrLinearSize);
|
2006-04-24 01:14:18 +02:00
|
|
|
if(sd_flags & DDSD_PITCH)
|
|
|
|
{
|
2008-05-01 17:41:42 +02:00
|
|
|
if(sd_pitch_or_size != round_up(pitch, 4u))
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::CORRUPTED);
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
if(sd_flags & DDSD_LINEARSIZE)
|
|
|
|
{
|
2006-12-15 17:09:30 +01:00
|
|
|
if(sd_pitch_or_size != pitch*stored_h)
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::CORRUPTED);
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
// note: both flags set would be invalid; no need to check for that,
|
|
|
|
// though, since one of the above tests would fail.
|
|
|
|
|
|
|
|
// mipmaps
|
|
|
|
if(sd_flags & DDSD_MIPMAPCOUNT)
|
|
|
|
{
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t mipmap_count = (size_t)read_le32(&sd->dwMipMapCount);
|
2006-04-24 01:14:18 +02:00
|
|
|
if(mipmap_count)
|
|
|
|
{
|
|
|
|
// mipmap chain is incomplete
|
|
|
|
// note: DDS includes the base level in its count, hence +1.
|
2007-05-09 23:01:11 +02:00
|
|
|
if(mipmap_count != ceil_log2(std::max(w,h))+1)
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::TEX_FMT_INVALID);
|
2006-04-24 01:14:18 +02:00
|
|
|
flags |= TEX_MIPMAPS;
|
|
|
|
}
|
|
|
|
}
|
|
|
|
|
|
|
|
// check for volume textures
|
|
|
|
if(sd_flags & DDSD_DEPTH)
|
|
|
|
{
|
2008-05-01 17:41:42 +02:00
|
|
|
const size_t depth = (size_t)read_le32(&sd->dwDepth);
|
2006-04-24 01:14:18 +02:00
|
|
|
if(depth)
|
2006-09-22 15:19:40 +02:00
|
|
|
WARN_RETURN(ERR::NOT_IMPLEMENTED);
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
// check caps
|
|
|
|
const DDCAPS2* caps = &sd->ddsCaps;
|
|
|
|
// .. this is supposed to be set, but don't bail if not (pointless)
|
|
|
|
debug_assert(caps->dwCaps1 & DDSCAPS_TEXTURE);
|
|
|
|
// .. sanity check: warn if mipmap flag not set (don't bail if not
|
|
|
|
// because we've already made the decision).
|
|
|
|
const bool mipmap_cap = (caps->dwCaps1 & DDSCAPS_MIPMAP) != 0;
|
|
|
|
const bool mipmap_flag = (flags & TEX_MIPMAPS) != 0;
|
|
|
|
debug_assert(mipmap_cap == mipmap_flag);
|
|
|
|
// note: we do not check for cubemaps and volume textures (not supported)
|
|
|
|
// because the file may still have useful data we can read.
|
|
|
|
|
|
|
|
*w_ = w;
|
|
|
|
*h_ = h;
|
|
|
|
*bpp_ = bpp;
|
|
|
|
*flags_ = flags;
|
2006-09-22 15:19:40 +02:00
|
|
|
return INFO::OK;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
//-----------------------------------------------------------------------------
|
|
|
|
|
|
|
|
static bool dds_is_hdr(const u8* file)
|
|
|
|
{
|
|
|
|
return *(u32*)file == FOURCC('D','D','S',' ');
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-12-22 19:15:52 +01:00
|
|
|
static bool dds_is_ext(const std::string& extension)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
2007-12-22 19:15:52 +01:00
|
|
|
return !strcasecmp(extension.c_str(), ".dds");
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
static size_t dds_hdr_size(const u8* UNUSED(file))
|
|
|
|
{
|
|
|
|
return 4+sizeof(DDSURFACEDESC2);
|
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-01-06 02:33:13 +01:00
|
|
|
static LibError dds_decode(DynArray* RESTRICT da, Tex* RESTRICT t)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
|
|
|
u8* file = da->base;
|
|
|
|
const DDSURFACEDESC2* sd = (const DDSURFACEDESC2*)(file+4);
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t w, h;
|
|
|
|
size_t bpp, flags;
|
2006-04-24 01:14:18 +02:00
|
|
|
RETURN_ERR(decode_sd(sd, &w, &h, &bpp, &flags));
|
|
|
|
// note: cannot pass address of these directly to decode_sd because
|
|
|
|
// they are bitfields.
|
|
|
|
t->w = w;
|
|
|
|
t->h = h;
|
|
|
|
t->bpp = bpp;
|
|
|
|
t->flags = flags;
|
2006-09-22 15:19:40 +02:00
|
|
|
return INFO::OK;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
2007-01-06 02:33:13 +01:00
|
|
|
static LibError dds_encode(Tex* RESTRICT UNUSED(t), DynArray* RESTRICT UNUSED(da))
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
2006-09-22 15:19:40 +02:00
|
|
|
// note: do not return ERR::NOT_IMPLEMENTED et al. because that would
|
2006-04-24 01:14:18 +02:00
|
|
|
// break tex_write (which assumes either this, 0 or errors are returned).
|
2006-09-22 15:19:40 +02:00
|
|
|
return INFO::TEX_CODEC_CANNOT_HANDLE;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
static LibError dds_transform(Tex* t, size_t transforms)
|
2006-04-24 01:14:18 +02:00
|
|
|
{
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
size_t dxt = t->flags & TEX_DXT;
|
2006-04-24 01:14:18 +02:00
|
|
|
debug_assert(is_valid_dxt(dxt));
|
|
|
|
|
had to remove uint and ulong from lib/types.h due to conflict with other library.
this snowballed into a massive search+destroy of the hodgepodge of
mostly equivalent types we had in use (int, uint, unsigned, unsigned
int, i32, u32, ulong, uintN).
it is more efficient to use 64-bit types in 64-bit mode, so the
preferred default is size_t (for anything remotely resembling a size or
index). tile coordinates are ssize_t to allow more efficient conversion
to/from floating point. flags are int because we almost never need more
than 15 distinct bits, bit test/set is not slower and int is fastest to
type. finally, some data that is pretty much directly passed to OpenGL
is now typed accordingly.
after several hours, the code now requires fewer casts and less
guesswork.
other changes:
- unit and player IDs now have an "invalid id" constant in the
respective class to avoid casting and -1
- fix some endian/64-bit bugs in the map (un)packing. added a
convenience function to write/read a size_t.
- ia32: change CPUID interface to allow passing in ecx (required for
cache topology detection, which I need at work). remove some unneeded
functions from asm, replace with intrinsics where possible.
This was SVN commit r5942.
2008-05-11 20:48:32 +02:00
|
|
|
const size_t transform_dxt = transforms & TEX_DXT;
|
2006-04-24 01:14:18 +02:00
|
|
|
// requesting decompression
|
|
|
|
if(dxt && transform_dxt)
|
|
|
|
{
|
|
|
|
RETURN_ERR(s3tc_decompress(t));
|
2006-09-22 15:19:40 +02:00
|
|
|
return INFO::OK;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
// both are DXT (unsupported; there are no flags we can change while
|
|
|
|
// compressed) or requesting compression (not implemented) or
|
|
|
|
// both not DXT (nothing we can do) - bail.
|
|
|
|
else
|
2006-09-22 15:19:40 +02:00
|
|
|
return INFO::TEX_CODEC_CANNOT_HANDLE;
|
2006-04-24 01:14:18 +02:00
|
|
|
}
|
|
|
|
|
|
|
|
|
|
|
|
TEX_CODEC_REGISTER(dds);
|