Make the floating point regression tests optional. It has been known for quite some time that these tests do not always generate the same results unless there is full SIMD coverage of the floating point algorithms in libjpeg-turbo. Further research reveals that there are basically three expected results: the results from our SSE SIMD extensions (which are slightly more accurate than the C code), results from the C code when running on a 32-bit FPU (or when using SSE instructions on an x86-64 CPU, which is the default with GCC), and results from the C code when running on a 64-bit FPU (which presumably uses double-precision arithmetic by default.) There is basically no way to determine which type of math will be used prior to run time, so it's best to just let the developers specify which result they expect on their particular system.

git-svn-id: svn+ssh://svn.code.sf.net/p/libjpeg-turbo/code/branches/1.4.x@1515 632fc199-4ca6-4c93-a231-07263d6284db
2015-01-16 06:53:36 +00:00
parent dd7f03be56
commit 3af282df3c
3 changed files with 60 additions and 13 deletions
--- a/ChangeLog.txt
+++ b/ChangeLog.txt
@@ -15,6 +15,13 @@ requires a color management system.
 mainly for the purpose of testing compression from/decompression to a subregion
 of a larger image buffer.

+[3] 'make test' no longer tests the regression of the floating point DCT/IDCT
+by default, since the results of those tests can vary if the algorithms in
+question are not implemented using SIMD instructions on a particular platform.
+See the comments in Makefile.am for information on how to re-enable the tests
+and to specify an expected result for them based on the particulars of your
+platform.
+

 1.4.0
 =====
--- a/Makefile.am
+++ b/Makefile.am
@@ -192,8 +192,13 @@ MD5_JPEG_GRAY_ISLOW = 235c90707b16e2e069f37c888b2636d9
 MD5_PPM_GRAY_ISLOW = 7213c10af507ad467da5578ca5ee1fca
 MD5_PPM_GRAY_ISLOW_RGB = e96ee81c30a6ed422d466338bd3de65d
 MD5_JPEG_420S_IFAST_OPT = 7af8e60be4d9c227ec63ac9b6630855e
-MD5_JPEG_3x2_FLOAT_PROG = a8c17daf77b457725ec929e215b603f8
-MD5_PPM_3x2_FLOAT = 42876ab9e5c2f76a87d08db5fbd57956
+MD5_JPEG_3x2_FLOAT_PROG_SSE = a8c17daf77b457725ec929e215b603f8
+MD5_PPM_3x2_FLOAT_SSE = 42876ab9e5c2f76a87d08db5fbd57956
+MD5_JPEG_3x2_FLOAT_PROG_32BIT = a8c17daf77b457725ec929e215b603f8
+MD5_PPM_3x2_FLOAT_32BIT = 42876ab9e5c2f76a87d08db5fbd57956
+MD5_PPM_3x2_FLOAT_64BIT = d6fbc71153b3d8ded484dbc17c7b9cf4
+MD5_JPEG_3x2_IFAST_PROG = 1396cc2b7185cfe943d408c9d305339e
+MD5_PPM_3x2_IFAST = 3975985ef6eeb0a2cdc58daa651ccc00
 MD5_PPM_420M_ISLOW_2_1 = 4ca6be2a6f326ff9eaab63e70a8259c0
 MD5_PPM_420M_ISLOW_15_8 = 12aa9f9534c1b3d7ba047322226365eb
 MD5_PPM_420M_ISLOW_13_8 = f7e22817c7b25e1393e4ec101e9d4e96
@@ -230,13 +235,13 @@ MD5_BMP_GRAY_ISLOW_565 = 12f78118e56a2f48b966f792fedf23cc
 MD5_BMP_GRAY_ISLOW_565D = bdbbd616441a24354c98553df5dc82db
 MD5_JPEG_420S_IFAST_OPT = 388708217ac46273ca33086b22827ed8
 # See README-turbo.txt for more details on why this next bit is necessary.
-if WITH_SSE_FLOAT_DCT
-MD5_JPEG_3x2_FLOAT_PROG = 343e3f8caf8af5986ebaf0bdc13b5c71
-MD5_PPM_3x2_FLOAT = 1a75f36e5904d6fc3a85a43da9ad89bb
-else
-MD5_JPEG_3x2_FLOAT_PROG = 9bca803d2042bd1eb03819e2bf92b3e5
-MD5_PPM_3x2_FLOAT = f6bfab038438ed8f5522fbd33595dcdc
-endif
+MD5_JPEG_3x2_FLOAT_PROG_SSE = 343e3f8caf8af5986ebaf0bdc13b5c71
+MD5_PPM_3x2_FLOAT_SSE = 1a75f36e5904d6fc3a85a43da9ad89bb
+MD5_JPEG_3x2_FLOAT_PROG_32BIT = 9bca803d2042bd1eb03819e2bf92b3e5
+MD5_PPM_3x2_FLOAT_32BIT = f6bfab038438ed8f5522fbd33595dcdc
+MD5_PPM_3x2_FLOAT_64BIT = 0e917a34193ef976b679a6b069b1be26
+MD5_JPEG_3x2_IFAST_PROG = 1ee5d2c1a77f2da495f993c8c7cceca5
+MD5_PPM_3x2_IFAST = fd283664b3b49127984af0a7f118fccd
 MD5_JPEG_420_ISLOW_ARI = e986fb0a637a8d833d96e8a6d6d84ea1
 MD5_JPEG_444_ISLOW_PROGARI = 0a8f1c8f66e113c3cf635df0a475a617
 MD5_PPM_420M_IFAST_ARI = 72b59a99bcf1de24c5b27d151bde2437
@@ -433,14 +438,46 @@ endif
 	md5/md5cmp $(MD5_JPEG_420S_IFAST_OPT) testout_420s_ifast_opt.jpg
 	rm testout_420s_ifast_opt.jpg

+# The output of the floating point tests is not validated by default, because
+# the output differs depending on the type of floating point math used, and
+# this is only deterministic if there is full SIMD coverage of all of the
+# floating point algorithms in libjpeg-turbo.  Pass one of the following on the
+# make command line to validate the floating point tests against one of the
+# expected results:
+#
+# FLOATTEST=sse  validate against the expected results from the libjpeg-turbo
+#                SSE SIMD extensions
+# FLOATTEST=32bit  validate against the expected results from the MIPS DSPr2
+#                  SIMD extensions or 32-bit FPUs or GCC when -mfpmath=sse is
+#                  used (which is the default on x86-64 systems)
+# FLOATTEST=64bit  validate against the exepected results from 64-bit FPUs
+
 # CC: RGB->YCC  SAMP: fullsize/int  FDCT: float  ENT: prog huff
 	./cjpeg -sample 3x2 -dct float -prog -outfile testout_3x2_float_prog.jpg $(srcdir)/testimages/testorig.ppm
-	md5/md5cmp $(MD5_JPEG_3x2_FLOAT_PROG) testout_3x2_float_prog.jpg
+	if [ "${FLOATTEST}" = "sse" ]; then \
+		md5/md5cmp $(MD5_JPEG_3x2_FLOAT_PROG_SSE) testout_3x2_float_prog.jpg; \
+	elif [ "${FLOATTEST}" = "32bit" -o "${FLOATTEST}" = "64bit" ]; then \
+		md5/md5cmp $(MD5_JPEG_3x2_FLOAT_PROG_32BIT) testout_3x2_float_prog.jpg; \
+	fi
 # CC: YCC->RGB  SAMP: fullsize/int  IDCT: float  ENT: prog huff
 	./djpeg -dct float -outfile testout_3x2_float.ppm testout_3x2_float_prog.jpg
-	md5/md5cmp $(MD5_PPM_3x2_FLOAT) testout_3x2_float.ppm
+	if [ "${FLOATTEST}" = "sse" ]; then \
+		md5/md5cmp $(MD5_PPM_3x2_FLOAT_SSE) testout_3x2_float.ppm; \
+	elif [ "${FLOATTEST}" = "32bit" ]; then \
+		md5/md5cmp $(MD5_PPM_3x2_FLOAT_32BIT) testout_3x2_float.ppm; \
+	elif [ "${FLOATTEST}" = "64bit" ]; then \
+		md5/md5cmp $(MD5_PPM_3x2_FLOAT_64BIT) testout_3x2_float.ppm; \
+	fi
 	rm testout_3x2_float.ppm testout_3x2_float_prog.jpg

+# CC: RGB->YCC  SAMP: fullsize/int  FDCT: ifast  ENT: prog huff
+	./cjpeg -sample 3x2 -dct fast -prog -outfile testout_3x2_ifast_prog.jpg $(srcdir)/testimages/testorig.ppm
+	md5/md5cmp $(MD5_JPEG_3x2_IFAST_PROG) testout_3x2_ifast_prog.jpg
+# CC: YCC->RGB  SAMP: fullsize/int  IDCT: ifast  ENT: prog huff
+	./djpeg -dct fast -outfile testout_3x2_ifast.ppm testout_3x2_ifast_prog.jpg
+	md5/md5cmp $(MD5_PPM_3x2_IFAST) testout_3x2_ifast.ppm
+	rm testout_3x2_ifast.ppm testout_3x2_ifast_prog.jpg
+
 if WITH_ARITH_ENC
 # CC: YCC->RGB  SAMP: fullsize/h2v2  FDCT: islow  ENT: arith
 	./cjpeg -dct int -arithmetic -outfile testout_420_islow_ari.jpg $(srcdir)/testimages/testorig.ppm
--- a/README-turbo.txt
+++ b/README-turbo.txt
@@ -311,8 +311,11 @@ following reasons:
   numbers on this, the typical difference in PNSR between the two algorithms
   is less than 0.10 dB, whereas changing the quality level by 1 in the upper
   range of the quality scale is typically more like a 1.0 dB difference.)
-- When not using the SIMD extensions, then the accuracy of the floating point
-   DCT/IDCT can depend on the compiler and compiler settings.
+-- If the floating point algorithms in libjpeg-turbo are not implemented using
+   SIMD instructions on a particular platform, then the accuracy of the
+   floating point DCT/IDCT can depend on the compiler settings.  For instance,
+   different results will be obtained when using -mfpmath=387 or -mfpmath=sse
+   with GCC on x86 systems.

 While libjpeg-turbo does emulate the libjpeg v8 API/ABI, under the hood, it is
 still using the same algorithms as libjpeg v6b, so there are several specific