The SIMD quantization algorithm does not produce correct results with the fast forward integer DCT and JPEG qualities >= 98, so for now, use the non-SIMD quantization function under those circumstances.

git-svn-id: https://libjpeg-turbo.svn.sourceforge.net/svnroot/libjpeg-turbo@383 3789f03b-4d11-0410-bbf8-ca57d06f2519

6 files changed, 61 insertions, 11 deletions

diff --git a/branches/1.0.x/ChangeLog.txt b/branches/1.0.x/ChangeLog.txt
index a1ad2e4..9b1c317 100644
--- a/branches/1.0.x/ChangeLog.txt
+++ b/branches/1.0.x/ChangeLog.txt
@@ -3,13 +3,21 @@
 
 [1] Added further protections against invalid Huffman codes.
 
-[2] Since the fast integer forward DCT seems to degrade for JPEG qualities
-greater than 95, TurboJPEG/OSS will now automatically use the slow integer
-forward DCT when generating JPEG images of quality 96 or greater.  This
-reduces compression performance by as much as 15% for these high-quality images
-but is necessary to ensure that the images are perceptually lossless.
-
-[3] Fixed visual artifacts in grayscale JPEG compression caused by a typo in
+[2] The algorithm used by the SIMD quantization function cannot produce correct
+results when the JPEG quality is >= 98 and the fast integer forward DCT is
+used.  Thus, the non-SIMD quantization function is now used for those cases,
+and libjpeg-turbo should now produce identical output to libjpeg v6b in all
+cases.
+
+[3] Despite the above, the fast integer forward DCT still degrades somewhat for
+JPEG qualities greater than 95, so TurboJPEG/OSS will now automatically use the
+slow integer forward DCT when generating JPEG images of quality 96 or greater.
+This reduces compression performance by as much as 15% for these high-quality
+images but is necessary to ensure that the images are perceptually lossless.
+It also ensures that the library can avoid the performance pitfall created by
+[2].
+
+[4] Fixed visual artifacts in grayscale JPEG compression caused by a typo in
 the RGB-to-chrominance lookup tables.
 
 
diff --git a/branches/1.0.x/Makefile.am b/branches/1.0.x/Makefile.am
index f8552ea..a22ecda 100644
--- a/branches/1.0.x/Makefile.am
+++ b/branches/1.0.x/Makefile.am
@@ -102,6 +102,8 @@ test: testclean all
 	./jpegut
 	./cjpeg -dct int -outfile testoutint.jpg $(srcdir)/testorig.ppm
 	./cjpeg -dct fast -opt -outfile testoutfst.jpg $(srcdir)/testorig.ppm
+	./cjpeg -dct fast -quality 100 -opt -outfile testoutfst100.jpg $(srcdir)/testorig.ppm
+	cmp $(srcdir)/testimgfst100.jpg testoutfst100.jpg
 	./cjpeg -dct float -outfile testoutflt.jpg $(srcdir)/testorig.ppm
 	cmp $(srcdir)/testimgint.jpg testoutint.jpg
 	cmp $(srcdir)/testimgfst.jpg testoutfst.jpg
diff --git a/branches/1.0.x/README-turbo.txt b/branches/1.0.x/README-turbo.txt
index 40350d0..8fb002b 100755
--- a/branches/1.0.x/README-turbo.txt
+++ b/branches/1.0.x/README-turbo.txt
@@ -43,6 +43,34 @@ counterparts from the libjpeg v6b source.
 
 
 *******************************************************************************
+**     Performance pitfalls
+*******************************************************************************
+
+===============
+Restart Markers
+===============
+
+The optimized Huffman decoder in libjpeg-turbo does not handle restart markers
+in a way that makes libjpeg happy, so it is necessary to use the slow Huffman
+decoder when decompressing a JPEG image that has restart markers.  This can
+cause the decompression performance to drop by as much as 20%, but the
+performance will still be much much greater than that of libjpeg v6b.  Many
+consumer packages, such as PhotoShop, use restart markers when generating JPEG
+images, so images generated by those programs will experience this issue.
+
+===============================================
+Fast Integer Forward DCT at High Quality Levels
+===============================================
+
+The algorithm used by the SIMD-accelerated quantization function cannot produce
+correct results whenever the fast integer forward DCT is used along with a JPEG
+quality of 98-100.  Thus, libjpeg-turbo must use the non-SIMD quantization
+function in those cases.  This causes performance to drop by as much as 40%.
+It is therefore strongly advised that you use the slow integer forward DCT
+whenever encoding images with a JPEG quality of 98 or higher.
+
+
+*******************************************************************************
 **     Using libjpeg-turbo
 *******************************************************************************
 
diff --git a/branches/1.0.x/jcdctmgr.c b/branches/1.0.x/jcdctmgr.c
index 156957a..711f9da 100644
--- a/branches/1.0.x/jcdctmgr.c
+++ b/branches/1.0.x/jcdctmgr.c
@@ -4,6 +4,7 @@
  * Copyright (C) 1994-1996, Thomas G. Lane.
  * Copyright (C) 1999-2006, MIYASAKA Masaru.
  * Copyright 2009 Pierre Ossman <ossman@cendio.se> for Cendio AB
+ * Copyright (C) 2011 D. R. Commander
  * This file is part of the Independent JPEG Group's software.
  * For conditions of distribution and use, see the accompanying README file.
  *
@@ -39,6 +40,8 @@ typedef JMETHOD(void, float_quantize_method_ptr,
                 (JCOEFPTR coef_block, FAST_FLOAT * divisors,
                  FAST_FLOAT * workspace));
 
+METHODDEF(void) quantize (JCOEFPTR, DCTELEM *, DCTELEM *);
+
 typedef struct {
   struct jpeg_forward_dct pub;	/* public fields */
 
@@ -160,7 +163,7 @@ flss (UINT16 val)
  * of in a consecutive manner, yet again in order to allow SIMD
  * routines.
  */
-LOCAL(void)
+LOCAL(int)
 compute_reciprocal (UINT16 divisor, DCTELEM * dtbl)
 {
   UDCTELEM2 fq, fr;
@@ -189,6 +192,9 @@ compute_reciprocal (UINT16 divisor, DCTELEM * dtbl)
   dtbl[DCTSIZE2 * 1] = (DCTELEM) c;       /* correction + roundfactor */
   dtbl[DCTSIZE2 * 2] = (DCTELEM) (1 << (sizeof(DCTELEM)*8*2 - r));  /* scale */
   dtbl[DCTSIZE2 * 3] = (DCTELEM) r - sizeof(DCTELEM)*8; /* shift */
+
+  if(r <= 16) return 0;
+  else return 1;
 }
 
 /*
@@ -232,7 +238,9 @@ start_pass_fdctmgr (j_compress_ptr cinfo)
       }
       dtbl = fdct->divisors[qtblno];
       for (i = 0; i < DCTSIZE2; i++) {
-	compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i]);
+	if(!compute_reciprocal(qtbl->quantval[i] << 3, &dtbl[i])
+	  && fdct->quantize == jsimd_quantize)
+	  fdct->quantize = quantize;
       }
       break;
 #endif
@@ -266,10 +274,12 @@ start_pass_fdctmgr (j_compress_ptr cinfo)
 	}
 	dtbl = fdct->divisors[qtblno];
 	for (i = 0; i < DCTSIZE2; i++) {
-	  compute_reciprocal(
+	  if(!compute_reciprocal(
 	    DESCALE(MULTIPLY16V16((INT32) qtbl->quantval[i],
 				  (INT32) aanscales[i]),
-		    CONST_BITS-3), &dtbl[i]);
+		    CONST_BITS-3), &dtbl[i])
+	    && fdct->quantize == jsimd_quantize)
+	    fdct->quantize = quantize;
 	}
       }
       break;
diff --git a/branches/1.0.x/testimgfst100.jpg b/branches/1.0.x/testimgfst100.jpg
new file mode 100644
index 0000000..36d9b75
--- /dev/null
+++ b/branches/1.0.x/testimgfst100.jpg
diff --git a/branches/1.0.x/win/Makefile b/branches/1.0.x/win/Makefile
index 9101c79..a63faf5 100755
--- a/branches/1.0.x/win/Makefile
+++ b/branches/1.0.x/win/Makefile
@@ -248,9 +248,11 @@ test: testclean $(ODIR)/cjpeg.exe $(ODIR)/djpeg.exe $(ODIR)/jpegtran.exe \
 	cd $(ODIR);	./jpegut
 	$(ODIR)/cjpeg -dct int -outfile $(ODIR)/testoutint.jpg testorig.ppm
 	$(ODIR)/cjpeg -dct fast -opt -outfile $(ODIR)/testoutfst.jpg testorig.ppm
+	$(ODIR)/cjpeg -dct fast -quality 100 -opt -outfile $(ODIR)/testoutfst100.jpg testorig.ppm
 	$(ODIR)/cjpeg -dct float -outfile $(ODIR)/testoutflt.jpg testorig.ppm
 	cmp testimgint.jpg $(ODIR)/testoutint.jpg
 	cmp testimgfst.jpg $(ODIR)/testoutfst.jpg
+	cmp testimgfst100.jpg $(ODIR)/testoutfst100.jpg
 	cmp testimgflt.jpg $(ODIR)/testoutflt.jpg
 	$(ODIR)/djpeg -dct int -fast -ppm -outfile $(ODIR)/testoutint.ppm testorig.jpg
 	$(ODIR)/djpeg -dct fast -ppm -outfile $(ODIR)/testoutfst.ppm testorig.jpg