#243 Regession in test pngvalid-transform on big endian between 1.6.19 and 1.6.20

[Antoine Brodin]

We see a regression between version 1.6.19 and version 1.6.20 on
FreeBSD/Sparc64 (big endian):

With version 1.6.19 all tests were succeeding.
With version 1.6.20:

============================================================================
Testsuite summary for libpng 1.6.20
============================================================================
# TOTAL: 32
# PASS:  31
# SKIP:  0
# XFAIL: 0
# FAIL:  1
# XPASS: 0
# ERROR: 0

FAIL: tests/pngvalid-transform
==============================

pngvalid: read: truecolour with alpha 8 bit: transform:
+rgb_to_gray^0.34483[overridden]: rgb_to_gray error 0.166824 exceeds
limit 0.165976
pngvalid: 1 errors, 0 warnings
FAIL: pngvalid --strict --transform (floating point arithmetic)
FAIL tests/pngvalid-transform (exit status: 1)

We also see this failure on mips and powerpc (big endian too).

[Glenn Randers-Pehrson]

Thanks, we will have a look. Can you provide more specific information about what compiler(s) you used and your hardware? Ths solution will probably be simple, just
increasing the tolerance, but we need to understand why it's necessary.

[Glenn Randers-Pehrson]

I was able to replicate the behavior on a Sun server:
SunOS hostname 5.10 Generic_144488-12 sun4u sparc SUNW,Sun-Blade-2500

libpng-1.6.19/tests/pngvalid-transform passes
lbipng-1.6.20/tests/pngvalid-transform fails

[Antoine Brodin]

This is on a SunFive V120 running FreeBSD 10.2 stable, with gcc 4.2.1

[Glenn Randers-Pehrson]

Deleted my previous comment saying the bug is fixed; it was about a different bug.

[Glenn Randers-Pehrson]

Please try libpng-1.6.21beta03. The pngvalid-transform test should be passing now.

[Antoine Brodin]

I confirm the test now pass on sparc64.

[Glenn Randers-Pehrson]

pngvalid-transform test was fixed in libpng-1.6.21. Thanks.