From 3adc64a4cb8655f4a44abd43b22fce0c09065e5d Mon Sep 17 00:00:00 2001 From: Frank Bossen Date: Wed, 23 Jul 2014 10:26:46 -0400 Subject: [PATCH] Improve floating point DCT From libjpeg-turbo r1288 Port the more accurate (and slightly faster) floating point IDCT implementation from jpeg-8a and later. New research revealed that the SSE/SSE2 floating point IDCT implementation was actually more accurate than the jpeg-6b implementation, not less, which is why its mathematical results have always differed from those of the jpeg-6b implementation. This patch brings the accuracy of the C code in line with that of the SSE/SSE2 code. --- ChangeLog.txt | 8 +++++ README-turbo.txt | 15 ++++++--- jidctflt.c | 79 +++++++++++++++++++++++------------------------- 3 files changed, 56 insertions(+), 46 deletions(-) diff --git a/ChangeLog.txt b/ChangeLog.txt index f44e0bb5..2c1e628b 100644 --- a/ChangeLog.txt +++ b/ChangeLog.txt @@ -23,6 +23,14 @@ and is out of scope for a codec library. [5] The TurboJPEG API can now be used to compress JPEG images from YUV planar source images. +[7] Improved the accuracy and performance of the non-SIMD implementation of the +floating point inverse DCT (using code borrowed from libjpeg v8a and later.) +The accuracy of this implementation now matches the accuracy of the SSE/SSE2 +implementation. Note, however, that the floating point DCT/IDCT algorithms are +mainly a legacy feature. They generally do not produce significantly better +accuracy than the slow integer DCT/IDCT algorithms, and they are quite a bit +slower. + 1.3.1 ===== diff --git a/README-turbo.txt b/README-turbo.txt index a94ff972..65da622c 100755 --- a/README-turbo.txt +++ b/README-turbo.txt @@ -426,10 +426,19 @@ following reasons: slightly more accurate than the implementation in libjpeg v6b, but not by any amount perceptible to human vision (generally in the range of 0.01 to 0.08 dB gain in PNSR.) +-- When not using the SIMD extensions, libjpeg-turbo uses the more accurate + (and slightly faster) floating point IDCT algorithm introduced in libjpeg + v8a as opposed to the algorithm used in libjpeg v6b. It should be noted, + however, that this algorithm basically brings the accuracy of the floating + point IDCT in line with the accuracy of the slow integer IDCT. The floating + point DCT/IDCT algorithms are mainly a legacy feature, and they do not + produce significantly more accuracy than the slow integer algorithms (to put + 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. - 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 cases in which libjpeg-turbo cannot be expected to produce the same output as @@ -445,10 +454,6 @@ libjpeg v8: output of libjpeg v8 is less accurate than that of libjpeg v6b for this reason. --- When using the floating point IDCT, for the reasons stated above and also - because the floating point IDCT algorithm was modified in libjpeg v8a to - improve accuracy. - -- When decompressing using a scaling factor > 1 and merged (AKA "non-fancy" or "non-smooth") chrominance upsampling, because libjpeg v8 does not support merged upsampling with scaling factors > 1. diff --git a/jidctflt.c b/jidctflt.c index 0188ce3d..f75b09b4 100644 --- a/jidctflt.c +++ b/jidctflt.c @@ -1,9 +1,12 @@ /* * jidctflt.c * + * This file was part of the Independent JPEG Group's software: * Copyright (C) 1994-1998, Thomas G. Lane. - * This file is part of the Independent JPEG Group's software. - * For conditions of distribution and use, see the accompanying README file. + * Modified 2010 by Guido Vollbeding. + * libjpeg-turbo Modifications: + * Copyright (C) 2014, D. R. Commander. + * For conditions of distribution and use, see the accompanying README file. * * This file contains a floating-point implementation of the * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine @@ -76,10 +79,10 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, FLOAT_MULT_TYPE * quantptr; FAST_FLOAT * wsptr; JSAMPROW outptr; - JSAMPLE *range_limit = IDCT_range_limit(cinfo); + JSAMPLE *range_limit = cinfo->sample_range_limit; int ctr; FAST_FLOAT workspace[DCTSIZE2]; /* buffers data between passes */ - SHIFT_TEMPS + #define _0_125 ((FLOAT_MULT_TYPE)0.125) /* Pass 1: process columns from input, store into work array. */ @@ -101,7 +104,8 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 && inptr[DCTSIZE*7] == 0) { /* AC terms all zero */ - FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]); + FAST_FLOAT dcval = DEQUANTIZE(inptr[DCTSIZE*0], + quantptr[DCTSIZE*0] * _0_125); wsptr[DCTSIZE*0] = dcval; wsptr[DCTSIZE*1] = dcval; @@ -120,10 +124,10 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, /* Even part */ - tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]); - tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]); - tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]); - tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]); + tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0] * _0_125); + tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2] * _0_125); + tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4] * _0_125); + tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6] * _0_125); tmp10 = tmp0 + tmp2; /* phase 3 */ tmp11 = tmp0 - tmp2; @@ -138,10 +142,10 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, /* Odd part */ - tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]); - tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]); - tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]); - tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]); + tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1] * _0_125); + tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3] * _0_125); + tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5] * _0_125); + tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7] * _0_125); z13 = tmp6 + tmp5; /* phase 6 */ z10 = tmp6 - tmp5; @@ -152,12 +156,12 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); /* 2*c4 */ z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */ - tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */ - tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */ + tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */ + tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */ tmp6 = tmp12 - tmp7; /* phase 2 */ tmp5 = tmp11 - tmp6; - tmp4 = tmp10 + tmp5; + tmp4 = tmp10 - tmp5; wsptr[DCTSIZE*0] = tmp0 + tmp7; wsptr[DCTSIZE*7] = tmp0 - tmp7; @@ -165,8 +169,8 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, wsptr[DCTSIZE*6] = tmp1 - tmp6; wsptr[DCTSIZE*2] = tmp2 + tmp5; wsptr[DCTSIZE*5] = tmp2 - tmp5; - wsptr[DCTSIZE*4] = tmp3 + tmp4; - wsptr[DCTSIZE*3] = tmp3 - tmp4; + wsptr[DCTSIZE*3] = tmp3 + tmp4; + wsptr[DCTSIZE*4] = tmp3 - tmp4; inptr++; /* advance pointers to next column */ quantptr++; @@ -174,7 +178,6 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, } /* Pass 2: process rows from work array, store into output array. */ - /* Note that we must descale the results by a factor of 8 == 2**3. */ wsptr = workspace; for (ctr = 0; ctr < DCTSIZE; ctr++) { @@ -187,8 +190,10 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, /* Even part */ - tmp10 = wsptr[0] + wsptr[4]; - tmp11 = wsptr[0] - wsptr[4]; + /* Apply signed->unsigned and prepare float->int conversion */ + z5 = wsptr[0] + ((FAST_FLOAT) CENTERJSAMPLE + (FAST_FLOAT) 0.5); + tmp10 = z5 + wsptr[4]; + tmp11 = z5 - wsptr[4]; tmp13 = wsptr[2] + wsptr[6]; tmp12 = (wsptr[2] - wsptr[6]) * ((FAST_FLOAT) 1.414213562) - tmp13; @@ -209,31 +214,23 @@ jpeg_idct_float (j_decompress_ptr cinfo, jpeg_component_info * compptr, tmp11 = (z11 - z13) * ((FAST_FLOAT) 1.414213562); z5 = (z10 + z12) * ((FAST_FLOAT) 1.847759065); /* 2*c2 */ - tmp10 = ((FAST_FLOAT) 1.082392200) * z12 - z5; /* 2*(c2-c6) */ - tmp12 = ((FAST_FLOAT) -2.613125930) * z10 + z5; /* -2*(c2+c6) */ + tmp10 = z5 - z12 * ((FAST_FLOAT) 1.082392200); /* 2*(c2-c6) */ + tmp12 = z5 - z10 * ((FAST_FLOAT) 2.613125930); /* 2*(c2+c6) */ tmp6 = tmp12 - tmp7; tmp5 = tmp11 - tmp6; - tmp4 = tmp10 + tmp5; + tmp4 = tmp10 - tmp5; - /* Final output stage: scale down by a factor of 8 and range-limit */ + /* Final output stage: float->int conversion and range-limit */ - outptr[0] = range_limit[(int) DESCALE((INT32) (tmp0 + tmp7), 3) - & RANGE_MASK]; - outptr[7] = range_limit[(int) DESCALE((INT32) (tmp0 - tmp7), 3) - & RANGE_MASK]; - outptr[1] = range_limit[(int) DESCALE((INT32) (tmp1 + tmp6), 3) - & RANGE_MASK]; - outptr[6] = range_limit[(int) DESCALE((INT32) (tmp1 - tmp6), 3) - & RANGE_MASK]; - outptr[2] = range_limit[(int) DESCALE((INT32) (tmp2 + tmp5), 3) - & RANGE_MASK]; - outptr[5] = range_limit[(int) DESCALE((INT32) (tmp2 - tmp5), 3) - & RANGE_MASK]; - outptr[4] = range_limit[(int) DESCALE((INT32) (tmp3 + tmp4), 3) - & RANGE_MASK]; - outptr[3] = range_limit[(int) DESCALE((INT32) (tmp3 - tmp4), 3) - & RANGE_MASK]; + outptr[0] = range_limit[((int) (tmp0 + tmp7)) & RANGE_MASK]; + outptr[7] = range_limit[((int) (tmp0 - tmp7)) & RANGE_MASK]; + outptr[1] = range_limit[((int) (tmp1 + tmp6)) & RANGE_MASK]; + outptr[6] = range_limit[((int) (tmp1 - tmp6)) & RANGE_MASK]; + outptr[2] = range_limit[((int) (tmp2 + tmp5)) & RANGE_MASK]; + outptr[5] = range_limit[((int) (tmp2 - tmp5)) & RANGE_MASK]; + outptr[3] = range_limit[((int) (tmp3 + tmp4)) & RANGE_MASK]; + outptr[4] = range_limit[((int) (tmp3 - tmp4)) & RANGE_MASK]; wsptr += DCTSIZE; /* advance pointer to next row */ }