• file
• revisions
• annotate
• diff
• comparison
• raw

media/libjpeg/jfdctfst.c@6474c204b198 (annotated)

media/libjpeg/jfdctfst.c

Wed, 31 Dec 2014 06:09:35 +0100

author

Michael Schloh von Bennewitz <michael@schloh.com>

date

Wed, 31 Dec 2014 06:09:35 +0100

changeset 0

6474c204b198

permissions

-rw-r--r--

Cloned upstream origin tor-browser at tor-browser-31.3.0esr-4.5-1-build1
revision ID fc1c9ff7c1b2defdbc039f12214767608f46423f for hacking purpose.

michael@0	1	/*
michael@0	2	* jfdctfst.c
michael@0	3	*
michael@0	4	* Copyright (C) 1994-1996, Thomas G. Lane.
michael@0	5	* This file is part of the Independent JPEG Group's software.
michael@0	6	* For conditions of distribution and use, see the accompanying README file.
michael@0	7	*
michael@0	8	* This file contains a fast, not so accurate integer implementation of the
michael@0	9	* forward DCT (Discrete Cosine Transform).
michael@0	10	*
michael@0	11	* A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT
michael@0	12	* on each column. Direct algorithms are also available, but they are
michael@0	13	* much more complex and seem not to be any faster when reduced to code.
michael@0	14	*
michael@0	15	* This implementation is based on Arai, Agui, and Nakajima's algorithm for
michael@0	16	* scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in
michael@0	17	* Japanese, but the algorithm is described in the Pennebaker & Mitchell
michael@0	18	* JPEG textbook (see REFERENCES section in file README). The following code
michael@0	19	* is based directly on figure 4-8 in P&M.
michael@0	20	* While an 8-point DCT cannot be done in less than 11 multiplies, it is
michael@0	21	* possible to arrange the computation so that many of the multiplies are
michael@0	22	* simple scalings of the final outputs. These multiplies can then be
michael@0	23	* folded into the multiplications or divisions by the JPEG quantization
michael@0	24	* table entries. The AA&N method leaves only 5 multiplies and 29 adds
michael@0	25	* to be done in the DCT itself.
michael@0	26	* The primary disadvantage of this method is that with fixed-point math,
michael@0	27	* accuracy is lost due to imprecise representation of the scaled
michael@0	28	* quantization values. The smaller the quantization table entry, the less
michael@0	29	* precise the scaled value, so this implementation does worse with high-
michael@0	30	* quality-setting files than with low-quality ones.
michael@0	31	*/
michael@0	32
michael@0	33	#define JPEG_INTERNALS
michael@0	34	#include "jinclude.h"
michael@0	35	#include "jpeglib.h"
michael@0	36	#include "jdct.h" /* Private declarations for DCT subsystem */
michael@0	37
michael@0	38	#ifdef DCT_IFAST_SUPPORTED
michael@0	39
michael@0	40
michael@0	41	/*
michael@0	42	* This module is specialized to the case DCTSIZE = 8.
michael@0	43	*/
michael@0	44
michael@0	45	#if DCTSIZE != 8
michael@0	46	Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
michael@0	47	#endif
michael@0	48
michael@0	49
michael@0	50	/* Scaling decisions are generally the same as in the LL&M algorithm;
michael@0	51	* see jfdctint.c for more details. However, we choose to descale
michael@0	52	* (right shift) multiplication products as soon as they are formed,
michael@0	53	* rather than carrying additional fractional bits into subsequent additions.
michael@0	54	* This compromises accuracy slightly, but it lets us save a few shifts.
michael@0	55	* More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
michael@0	56	* everywhere except in the multiplications proper; this saves a good deal
michael@0	57	* of work on 16-bit-int machines.
michael@0	58	*
michael@0	59	* Again to save a few shifts, the intermediate results between pass 1 and
michael@0	60	* pass 2 are not upscaled, but are represented only to integral precision.
michael@0	61	*
michael@0	62	* A final compromise is to represent the multiplicative constants to only
michael@0	63	* 8 fractional bits, rather than 13. This saves some shifting work on some
michael@0	64	* machines, and may also reduce the cost of multiplication (since there
michael@0	65	* are fewer one-bits in the constants).
michael@0	66	*/
michael@0	67
michael@0	68	#define CONST_BITS 8
michael@0	69
michael@0	70
michael@0	71	/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
michael@0	72	* causing a lot of useless floating-point operations at run time.
michael@0	73	* To get around this we use the following pre-calculated constants.
michael@0	74	* If you change CONST_BITS you may want to add appropriate values.
michael@0	75	* (With a reasonable C compiler, you can just rely on the FIX() macro...)
michael@0	76	*/
michael@0	77
michael@0	78	#if CONST_BITS == 8
michael@0	79	#define FIX_0_382683433 ((INT32) 98) /* FIX(0.382683433) */
michael@0	80	#define FIX_0_541196100 ((INT32) 139) /* FIX(0.541196100) */
michael@0	81	#define FIX_0_707106781 ((INT32) 181) /* FIX(0.707106781) */
michael@0	82	#define FIX_1_306562965 ((INT32) 334) /* FIX(1.306562965) */
michael@0	83	#else
michael@0	84	#define FIX_0_382683433 FIX(0.382683433)
michael@0	85	#define FIX_0_541196100 FIX(0.541196100)
michael@0	86	#define FIX_0_707106781 FIX(0.707106781)
michael@0	87	#define FIX_1_306562965 FIX(1.306562965)
michael@0	88	#endif
michael@0	89
michael@0	90
michael@0	91	/* We can gain a little more speed, with a further compromise in accuracy,
michael@0	92	* by omitting the addition in a descaling shift. This yields an incorrectly
michael@0	93	* rounded result half the time...
michael@0	94	*/
michael@0	95
michael@0	96	#ifndef USE_ACCURATE_ROUNDING
michael@0	97	#undef DESCALE
michael@0	98	#define DESCALE(x,n) RIGHT_SHIFT(x, n)
michael@0	99	#endif
michael@0	100
michael@0	101
michael@0	102	/* Multiply a DCTELEM variable by an INT32 constant, and immediately
michael@0	103	* descale to yield a DCTELEM result.
michael@0	104	*/
michael@0	105
michael@0	106	#define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
michael@0	107
michael@0	108
michael@0	109	/*
michael@0	110	* Perform the forward DCT on one block of samples.
michael@0	111	*/
michael@0	112
michael@0	113	GLOBAL(void)
michael@0	114	jpeg_fdct_ifast (DCTELEM * data)
michael@0	115	{
michael@0	116	DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
michael@0	117	DCTELEM tmp10, tmp11, tmp12, tmp13;
michael@0	118	DCTELEM z1, z2, z3, z4, z5, z11, z13;
michael@0	119	DCTELEM *dataptr;
michael@0	120	int ctr;
michael@0	121	SHIFT_TEMPS
michael@0	122
michael@0	123	/* Pass 1: process rows. */
michael@0	124
michael@0	125	dataptr = data;
michael@0	126	for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
michael@0	127	tmp0 = dataptr[0] + dataptr[7];
michael@0	128	tmp7 = dataptr[0] - dataptr[7];
michael@0	129	tmp1 = dataptr[1] + dataptr[6];
michael@0	130	tmp6 = dataptr[1] - dataptr[6];
michael@0	131	tmp2 = dataptr[2] + dataptr[5];
michael@0	132	tmp5 = dataptr[2] - dataptr[5];
michael@0	133	tmp3 = dataptr[3] + dataptr[4];
michael@0	134	tmp4 = dataptr[3] - dataptr[4];
michael@0	135
michael@0	136	/* Even part */
michael@0	137
michael@0	138	tmp10 = tmp0 + tmp3; /* phase 2 */
michael@0	139	tmp13 = tmp0 - tmp3;
michael@0	140	tmp11 = tmp1 + tmp2;
michael@0	141	tmp12 = tmp1 - tmp2;
michael@0	142
michael@0	143	dataptr[0] = tmp10 + tmp11; /* phase 3 */
michael@0	144	dataptr[4] = tmp10 - tmp11;
michael@0	145
michael@0	146	z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
michael@0	147	dataptr[2] = tmp13 + z1; /* phase 5 */
michael@0	148	dataptr[6] = tmp13 - z1;
michael@0	149
michael@0	150	/* Odd part */
michael@0	151
michael@0	152	tmp10 = tmp4 + tmp5; /* phase 2 */
michael@0	153	tmp11 = tmp5 + tmp6;
michael@0	154	tmp12 = tmp6 + tmp7;
michael@0	155
michael@0	156	/* The rotator is modified from fig 4-8 to avoid extra negations. */
michael@0	157	z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
michael@0	158	z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
michael@0	159	z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
michael@0	160	z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
michael@0	161
michael@0	162	z11 = tmp7 + z3; /* phase 5 */
michael@0	163	z13 = tmp7 - z3;
michael@0	164
michael@0	165	dataptr[5] = z13 + z2; /* phase 6 */
michael@0	166	dataptr[3] = z13 - z2;
michael@0	167	dataptr[1] = z11 + z4;
michael@0	168	dataptr[7] = z11 - z4;
michael@0	169
michael@0	170	dataptr += DCTSIZE; /* advance pointer to next row */
michael@0	171	}
michael@0	172
michael@0	173	/* Pass 2: process columns. */
michael@0	174
michael@0	175	dataptr = data;
michael@0	176	for (ctr = DCTSIZE-1; ctr >= 0; ctr--) {
michael@0	177	tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7];
michael@0	178	tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7];
michael@0	179	tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6];
michael@0	180	tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6];
michael@0	181	tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5];
michael@0	182	tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5];
michael@0	183	tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4];
michael@0	184	tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4];
michael@0	185
michael@0	186	/* Even part */
michael@0	187
michael@0	188	tmp10 = tmp0 + tmp3; /* phase 2 */
michael@0	189	tmp13 = tmp0 - tmp3;
michael@0	190	tmp11 = tmp1 + tmp2;
michael@0	191	tmp12 = tmp1 - tmp2;
michael@0	192
michael@0	193	dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */
michael@0	194	dataptr[DCTSIZE*4] = tmp10 - tmp11;
michael@0	195
michael@0	196	z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */
michael@0	197	dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */
michael@0	198	dataptr[DCTSIZE*6] = tmp13 - z1;
michael@0	199
michael@0	200	/* Odd part */
michael@0	201
michael@0	202	tmp10 = tmp4 + tmp5; /* phase 2 */
michael@0	203	tmp11 = tmp5 + tmp6;
michael@0	204	tmp12 = tmp6 + tmp7;
michael@0	205
michael@0	206	/* The rotator is modified from fig 4-8 to avoid extra negations. */
michael@0	207	z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */
michael@0	208	z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */
michael@0	209	z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */
michael@0	210	z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */
michael@0	211
michael@0	212	z11 = tmp7 + z3; /* phase 5 */
michael@0	213	z13 = tmp7 - z3;
michael@0	214
michael@0	215	dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */
michael@0	216	dataptr[DCTSIZE*3] = z13 - z2;
michael@0	217	dataptr[DCTSIZE*1] = z11 + z4;
michael@0	218	dataptr[DCTSIZE*7] = z11 - z4;
michael@0	219
michael@0	220	dataptr++; /* advance pointer to next column */
michael@0	221	}
michael@0	222	}
michael@0	223
michael@0	224	#endif /* DCT_IFAST_SUPPORTED */