Palette index checking: checking on read was erroneously skipped. pngcp can now
turn off the palette index checking or fix the errors (mainly to allow
comparison with libpng 1.6, which defaulted to ignoring it). The API now
documents how the 'enabled' parameter works. On read the check is on by default
(it was apparently off in libpng 1.6) however now if explicitly turned on the
warning message is stopped, this provides better interaction with
get_palette_max at the cost of a small API change (may need to be reviewed.)
Palette size errors: invalid entries in the palette are now set to #beaded to
make the errors more obvious and allow easy detection in memory.
Read transform pipeline caching: changes to fix the palette index check (which
are erroneously cached), the pipeline can now contain multiple caches. E.g.
caching of packing of palette indices can be combined with caching of palette
RGB transforms.
Read code now checks for callbacks to read 0 data (and faults them). Fixed the
reading of PNGs with multiple 0 length IDATs that failed as a result plus the
handling of zero length unknown. (Which occurs, validly; the spurious warning
has been removed).
filter selection: the 1.6 and earlier sum of absolute differences algorithm has
been reintroduced with an option to disfavor some filters over others where the
sums are close (not yet exposed). The selection code also logs the last known
occurence of each possible byte code across multiple lines. This allows
detection of PNG images with lower bit depth than the format implies and,
therefore, allows the filtering to be turned off in those cases as well.
The default write zlib settings are still lower than libpng16. The selection
algorithm is being refined (the overall results are worse than not doing any
filtering).
Signed-off-by: John Bowler <jbowler@acm.org>
Implemented better defaulting of zlib settings based on image properties.
Implemented pass-through of png_write_rows when the rows can be used directly (a
common case) optimizing the handling of previous-row buffering.
Removed the METHODICAL filter selection method and disabled the HEURISTIC one;
the first was ridiculously slow (though useful for experiments) the second
doesn't work. Filter selection is temporarily disabled (it defaults to the
lowest numbered filter in the list; typically 'none').
New handling of compression settings (incomplete), new PNG compression level
(not yet visible in an API).
Back ported 'PNG_FAST_FILTERS' from 1.6 (in png.h).
There are minimal API changes beyond removal of the selection options. Work is
still to be done to investigate a filter selection mechanism that is at least as
good as the previous one.
Signed-off-by: John Bowler <jbowler@acm.org>
Unify the compression code so that inflate calls are localized to a common
routine. Ground work for filter selection support. Minor API changes to use
void* not byte* for data parameters. Unification of some of the compression
code with the decompression code; IDAT_size replaces IDAT_read_size and
zbuffer_size, IDAT reading and writing is no longer controlled by the size of
the compression buffer.
Signed-off-by: John Bowler <jbowler@acm.org>
This implements a new chunk parse implementation that can be shared, it
is currently shared by the progressive reader and the sequential one
(not, yet, the writer).
The patch also implements shared transform handling that is used
throughout.
Signed-off-by: John Bowler <jbowler@acm.org>
Fixed 'minimal' builds. Various obviously useful minimal configurations
don't build because of missing contrib/libtests test programs and overly
complex dependencies in scripts/pnglibconf.dfa. This change adds
contrib/conftest/*.dfa files that can be used in automatic build
scripts to ensure that these configurations continue to build.
Enabled WRITE_INVERT and WRITE_PACK in contrib/pngminim/encoder.
png_malloc_default png_free_default.
Updated some left over "1.6.0beta32" in code sources.
Fixed a "png_structp" prototype (should be png_structrp) in arm_init.c
Updated the version-number hack in pngvalid.c
These changes cause 16-bit arithmetic to be used for 8-bit data in the gamma
corrected compose and grayscale operations. The arithmetic errors have
three sources all of which are fixed in this commit:
1) 8-bit linear calculations produce massive errors for lower intensity
values.
2) The old 16-bit "16 to 8" gamma table code erroneously wrote the lowest
output value into a table entry which corresponded to multiple output
values (so where the value written should have been the closest to the
transformed input value.)
3) In a number of cases the code to access the 16-bit table did not round;
it did a simple shift, which was wrong and made the side effects of (2)
even worse.
The new gamma code does not have the 16-to-8 problem at the cost of slighly
more calculations and the algorithm used to minimize the number of
calculations has been extended to all the 16-bit tables; it has advantages
for any significant gamma correction.
Added new "png_structrp" typedef. Because of the
way libpng works both png_info and png_struct are always accessed via a
single pointer. This means adding C99 'restrict' to the pointer gives
the compiler some opportunity to optimize the code. This change allows that.
changes alter how the tricky allocation of the initial png_struct and png_info
structures are handled. png_info is now handled in pretty much the same
way as everything else, except that the allocations handle NULL return
silently. png_struct is changed in a similar way on allocation and on
deallocation a 'safety' error handler is put in place (which should never
be required). The error handler itself is changed to permit mismatches
in the application and libpng error buffer size; however, this means a
silent change to the API to return the jmp_buf if the size doesn't match
the size from the libpng compilation; libpng now allocates the memory and
this may fail. Overall these changes result in slight code size
reductions; however, this is a reduction in code that is always executed
so is particularly valuable. Overall on a 64-bit system the libpng DLL
decreases in code size by 1733 bytes. pngerror.o increases in size by
about 465 bytes because of the new functionality.
literal pointers declare them (John Bowler).
Many APIs did not change their arguments but were not declared using
PNG_CONST. This change corrects this. In a few cases APIs that return
constant string literal pointers have also been changed to declare this.
Unlike the argument change this may require app changes; however the
results could never be written to (the app would crash on some platforms
where strings are not writable), so this seems advisable.
Fixed point APIs are now supported throughout (no missing APIs).
Internal fixed point arithmetic support exists for all internal floating
point operations.
sCAL validates the floating point strings it is passed.
Safe, albeit rudimentary, Watcom support is provided by PNG_API_RULE==2
Two new APIs exist to get the number of passes without turning on the
PNG_INTERLACE transform and to get the number of rows in the current
pass.
A new test program, pngvalid.c, validates the gamma code.
Errors in the 16 bit gamma correction (overflows) have been corrected.
cHRM chunk testing is done consistently (previously the floating point
API bypassed it, because the test really didn't work on FP, now the test
is performed on the actual values to be stored in the PNG file so it
works in the FP case too.)
Most floating point APIs now simply call the fixed point APIs after
converting the values to the fixed point form used in the PNG file.
The standard headers no longer include zlib.h, which is currently only
required for pngstruct.h and can therefore be internal.
(Patches by John Bowler)
