1e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#if !defined(_FX_JPEG_TURBO_)
2e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/*
3e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * jidctfst.c
4e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov *
5e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * Copyright (C) 1994-1998, Thomas G. Lane.
6e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * This file is part of the Independent JPEG Group's software.
7e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * For conditions of distribution and use, see the accompanying README file.
8e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov *
9e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * This file contains a fast, not so accurate integer implementation of the
10e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * inverse DCT (Discrete Cosine Transform).  In the IJG code, this routine
11e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * must also perform dequantization of the input coefficients.
12e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov *
13e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT
14e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * on each row (or vice versa, but it's more convenient to emit a row at
15e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * a time).  Direct algorithms are also available, but they are much more
16e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * complex and seem not to be any faster when reduced to code.
17e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov *
18e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * This implementation is based on Arai, Agui, and Nakajima's algorithm for
19e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * scaled DCT.  Their original paper (Trans. IEICE E-71(11):1095) is in
20e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * Japanese, but the algorithm is described in the Pennebaker & Mitchell
21e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * JPEG textbook (see REFERENCES section in file README).  The following code
22e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * is based directly on figure 4-8 in P&M.
23e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * While an 8-point DCT cannot be done in less than 11 multiplies, it is
24e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * possible to arrange the computation so that many of the multiplies are
25e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * simple scalings of the final outputs.  These multiplies can then be
26e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * folded into the multiplications or divisions by the JPEG quantization
27e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * table entries.  The AA&N method leaves only 5 multiplies and 29 adds
28e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * to be done in the DCT itself.
29e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * The primary disadvantage of this method is that with fixed-point math,
30e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * accuracy is lost due to imprecise representation of the scaled
31e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * quantization values.  The smaller the quantization table entry, the less
32e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * precise the scaled value, so this implementation does worse with high-
33e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * quality-setting files than with low-quality ones.
34e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
35e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
36e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define JPEG_INTERNALS
37e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#include "jinclude.h"
38e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#include "jpeglib.h"
39e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#include "jdct.h"		/* Private declarations for DCT subsystem */
40e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
41e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#ifdef DCT_IFAST_SUPPORTED
42e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
43e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
44e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/*
45e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * This module is specialized to the case DCTSIZE = 8.
46e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
47e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
48e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#if DCTSIZE != 8
49e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */
50e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
51e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
52e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
53e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/* Scaling decisions are generally the same as in the LL&M algorithm;
54e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * see jidctint.c for more details.  However, we choose to descale
55e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * (right shift) multiplication products as soon as they are formed,
56e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * rather than carrying additional fractional bits into subsequent additions.
57e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * This compromises accuracy slightly, but it lets us save a few shifts.
58e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples)
59e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * everywhere except in the multiplications proper; this saves a good deal
60e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * of work on 16-bit-int machines.
61e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov *
62e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * The dequantized coefficients are not integers because the AA&N scaling
63e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * factors have been incorporated.  We represent them scaled up by PASS1_BITS,
64e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * so that the first and second IDCT rounds have the same input scaling.
65e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to
66e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * avoid a descaling shift; this compromises accuracy rather drastically
67e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * for small quantization table entries, but it saves a lot of shifts.
68e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway,
69e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * so we use a much larger scaling factor to preserve accuracy.
70e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov *
71e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * A final compromise is to represent the multiplicative constants to only
72e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * 8 fractional bits, rather than 13.  This saves some shifting work on some
73e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * machines, and may also reduce the cost of multiplication (since there
74e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * are fewer one-bits in the constants).
75e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
76e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
77e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#if BITS_IN_JSAMPLE == 8
78e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define CONST_BITS  8
79e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define PASS1_BITS  2
80e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#else
81e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define CONST_BITS  8
82e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define PASS1_BITS  1		/* lose a little precision to avoid overflow */
83e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
84e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
85e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/* Some C compilers fail to reduce "FIX(constant)" at compile time, thus
86e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * causing a lot of useless floating-point operations at run time.
87e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * To get around this we use the following pre-calculated constants.
88e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * If you change CONST_BITS you may want to add appropriate values.
89e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * (With a reasonable C compiler, you can just rely on the FIX() macro...)
90e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
91e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
92e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#if CONST_BITS == 8
93e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_1_082392200  ((INT32)  277)		/* FIX(1.082392200) */
94e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_1_414213562  ((INT32)  362)		/* FIX(1.414213562) */
95e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_1_847759065  ((INT32)  473)		/* FIX(1.847759065) */
96e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_2_613125930  ((INT32)  669)		/* FIX(2.613125930) */
97e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#else
98e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_1_082392200  FIX(1.082392200)
99e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_1_414213562  FIX(1.414213562)
100e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_1_847759065  FIX(1.847759065)
101e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define FIX_2_613125930  FIX(2.613125930)
102e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
103e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
104e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
105e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/* We can gain a little more speed, with a further compromise in accuracy,
106e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * by omitting the addition in a descaling shift.  This yields an incorrectly
107e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * rounded result half the time...
108e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
109e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
110e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#ifndef USE_ACCURATE_ROUNDING
111e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#undef DESCALE
112e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define DESCALE(x,n)  RIGHT_SHIFT(x, n)
113e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
114e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
115e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
116e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/* Multiply a DCTELEM variable by an INT32 constant, and immediately
117e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * descale to yield a DCTELEM result.
118e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
119e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
120e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define MULTIPLY(var,const)  ((DCTELEM) DESCALE((var) * (const), CONST_BITS))
121e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
122e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
123e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/* Dequantize a coefficient by multiplying it by the multiplier-table
124e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * entry; produce a DCTELEM result.  For 8-bit data a 16x16->16
125e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * multiplication will do.  For 12-bit data, the multiplier table is
126e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * declared INT32, so a 32-bit multiply will be used.
127e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
128e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
129e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#if BITS_IN_JSAMPLE == 8
130e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define DEQUANTIZE(coef,quantval)  (((IFAST_MULT_TYPE) (coef)) * (quantval))
131e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#else
132e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define DEQUANTIZE(coef,quantval)  \
133e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov	DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS)
134e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
135e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
136e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
137e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/* Like DESCALE, but applies to a DCTELEM and produces an int.
138e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * We assume that int right shift is unsigned if INT32 right shift is.
139e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
140e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
141e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#ifdef RIGHT_SHIFT_IS_UNSIGNED
142e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define ISHIFT_TEMPS	DCTELEM ishift_temp;
143e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#if BITS_IN_JSAMPLE == 8
144e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define DCTELEMBITS  16		/* DCTELEM may be 16 or 32 bits */
145e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#else
146e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define DCTELEMBITS  32		/* DCTELEM must be 32 bits */
147e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
148e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define IRIGHT_SHIFT(x,shft)  \
149e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    ((ishift_temp = (x)) < 0 ? \
150e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \
151e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     (ishift_temp >> (shft)))
152e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#else
153e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define ISHIFT_TEMPS
154e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define IRIGHT_SHIFT(x,shft)	((x) >> (shft))
155e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
156e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
157e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#ifdef USE_ACCURATE_ROUNDING
158e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n))
159e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#else
160e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#define IDESCALE(x,n)  ((int) IRIGHT_SHIFT(x, n))
161e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
162e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
163e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
164e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov/*
165e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov * Perform dequantization and inverse DCT on one block of coefficients.
166e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov */
167e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
168e6986e1e8d4a57987f47c215490cb080a65ee29aSvet GanovGLOBAL(void)
169e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganovjpeg_idct_ifast (j_decompress_ptr cinfo, jpeg_component_info * compptr,
170e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov		 JCOEFPTR coef_block,
171e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov		 JSAMPARRAY output_buf, JDIMENSION output_col)
172e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov{
173e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
174e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  DCTELEM tmp10, tmp11, tmp12, tmp13;
175e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  DCTELEM z5, z10, z11, z12, z13;
176e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  JCOEFPTR inptr;
177e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  IFAST_MULT_TYPE * quantptr;
178e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  int * wsptr;
179e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  JSAMPROW outptr;
180e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  JSAMPLE *range_limit = IDCT_range_limit(cinfo);
181e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  int ctr;
182e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  int workspace[DCTSIZE2];	/* buffers data between passes */
183e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  SHIFT_TEMPS			/* for DESCALE */
184e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  ISHIFT_TEMPS			/* for IDESCALE */
185e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
186e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  /* Pass 1: process columns from input, store into work array. */
187e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
188e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  inptr = coef_block;
189e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  quantptr = (IFAST_MULT_TYPE *) compptr->dct_table;
190e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  wsptr = workspace;
191e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  for (ctr = DCTSIZE; ctr > 0; ctr--) {
192e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    /* Due to quantization, we will usually find that many of the input
193e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * coefficients are zero, especially the AC terms.  We can exploit this
194e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * by short-circuiting the IDCT calculation for any column in which all
195e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * the AC terms are zero.  In that case each output is equal to the
196e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * DC coefficient (with scale factor as needed).
197e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * With typical images and quantization tables, half or more of the
198e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * column DCT calculations can be simplified this way.
199e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     */
200e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
201e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    if (inptr[DCTSIZE*1] == 0 && inptr[DCTSIZE*2] == 0 &&
202e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov	inptr[DCTSIZE*3] == 0 && inptr[DCTSIZE*4] == 0 &&
203e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov	inptr[DCTSIZE*5] == 0 && inptr[DCTSIZE*6] == 0 &&
204e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov	inptr[DCTSIZE*7] == 0) {
205e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      /* AC terms all zero */
206e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      int dcval = (int) DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
207e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
208e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*0] = dcval;
209e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*1] = dcval;
210e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*2] = dcval;
211e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*3] = dcval;
212e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*4] = dcval;
213e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*5] = dcval;
214e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*6] = dcval;
215e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr[DCTSIZE*7] = dcval;
216e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
217e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      inptr++;			/* advance pointers to next column */
218e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      quantptr++;
219e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr++;
220e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      continue;
221e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    }
222e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
223e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    /* Even part */
224e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
225e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp0 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]);
226e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp1 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]);
227e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp2 = DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]);
228e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp3 = DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]);
229e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
230e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp10 = tmp0 + tmp2;	/* phase 3 */
231e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp11 = tmp0 - tmp2;
232e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
233e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp13 = tmp1 + tmp3;	/* phases 5-3 */
234e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp12 = MULTIPLY(tmp1 - tmp3, FIX_1_414213562) - tmp13; /* 2*c4 */
235e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
236e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp0 = tmp10 + tmp13;	/* phase 2 */
237e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp3 = tmp10 - tmp13;
238e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp1 = tmp11 + tmp12;
239e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp2 = tmp11 - tmp12;
240e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
241e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    /* Odd part */
242e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
243e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp4 = DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]);
244e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp5 = DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]);
245e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp6 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]);
246e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp7 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]);
247e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
248e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z13 = tmp6 + tmp5;		/* phase 6 */
249e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z10 = tmp6 - tmp5;
250e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z11 = tmp4 + tmp7;
251e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z12 = tmp4 - tmp7;
252e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
253e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp7 = z11 + z13;		/* phase 5 */
254e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
255e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
256e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
257e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
258e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
259e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
260e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp6 = tmp12 - tmp7;	/* phase 2 */
261e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp5 = tmp11 - tmp6;
262e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp4 = tmp10 + tmp5;
263e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
264e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*0] = (int) (tmp0 + tmp7);
265e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*7] = (int) (tmp0 - tmp7);
266e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*1] = (int) (tmp1 + tmp6);
267e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*6] = (int) (tmp1 - tmp6);
268e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*2] = (int) (tmp2 + tmp5);
269e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*5] = (int) (tmp2 - tmp5);
270e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*4] = (int) (tmp3 + tmp4);
271e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr[DCTSIZE*3] = (int) (tmp3 - tmp4);
272e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
273e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    inptr++;			/* advance pointers to next column */
274e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    quantptr++;
275e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr++;
276e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  }
277e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
278e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  /* Pass 2: process rows from work array, store into output array. */
279e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  /* Note that we must descale the results by a factor of 8 == 2**3, */
280e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  /* and also undo the PASS1_BITS scaling. */
281e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
282e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  wsptr = workspace;
283e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  for (ctr = 0; ctr < DCTSIZE; ctr++) {
284e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr = output_buf[ctr] + output_col;
285e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    /* Rows of zeroes can be exploited in the same way as we did with columns.
286e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * However, the column calculation has created many nonzero AC terms, so
287e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * the simplification applies less often (typically 5% to 10% of the time).
288e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * On machines with very fast multiplication, it's possible that the
289e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * test takes more time than it's worth.  In that case this section
290e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     * may be commented out.
291e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov     */
292e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
293e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#ifndef NO_ZERO_ROW_TEST
294e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    if (wsptr[1] == 0 && wsptr[2] == 0 && wsptr[3] == 0 && wsptr[4] == 0 &&
295e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov	wsptr[5] == 0 && wsptr[6] == 0 && wsptr[7] == 0) {
296e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      /* AC terms all zero */
297e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      JSAMPLE dcval = range_limit[IDESCALE(wsptr[0], PASS1_BITS+3)
298e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov				  & RANGE_MASK];
299e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
300e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[0] = dcval;
301e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[1] = dcval;
302e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[2] = dcval;
303e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[3] = dcval;
304e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[4] = dcval;
305e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[5] = dcval;
306e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[6] = dcval;
307e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      outptr[7] = dcval;
308e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
309e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      wsptr += DCTSIZE;		/* advance pointer to next row */
310e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov      continue;
311e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    }
312e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif
313e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
314e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    /* Even part */
315e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
316e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp10 = ((DCTELEM) wsptr[0] + (DCTELEM) wsptr[4]);
317e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp11 = ((DCTELEM) wsptr[0] - (DCTELEM) wsptr[4]);
318e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
319e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp13 = ((DCTELEM) wsptr[2] + (DCTELEM) wsptr[6]);
320e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp12 = MULTIPLY((DCTELEM) wsptr[2] - (DCTELEM) wsptr[6], FIX_1_414213562)
321e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov	    - tmp13;
322e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
323e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp0 = tmp10 + tmp13;
324e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp3 = tmp10 - tmp13;
325e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp1 = tmp11 + tmp12;
326e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp2 = tmp11 - tmp12;
327e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
328e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    /* Odd part */
329e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
330e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z13 = (DCTELEM) wsptr[5] + (DCTELEM) wsptr[3];
331e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z10 = (DCTELEM) wsptr[5] - (DCTELEM) wsptr[3];
332e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z11 = (DCTELEM) wsptr[1] + (DCTELEM) wsptr[7];
333e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z12 = (DCTELEM) wsptr[1] - (DCTELEM) wsptr[7];
334e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
335e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp7 = z11 + z13;		/* phase 5 */
336e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp11 = MULTIPLY(z11 - z13, FIX_1_414213562); /* 2*c4 */
337e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
338e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    z5 = MULTIPLY(z10 + z12, FIX_1_847759065); /* 2*c2 */
339e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp10 = MULTIPLY(z12, FIX_1_082392200) - z5; /* 2*(c2-c6) */
340e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp12 = MULTIPLY(z10, - FIX_2_613125930) + z5; /* -2*(c2+c6) */
341e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
342e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp6 = tmp12 - tmp7;	/* phase 2 */
343e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp5 = tmp11 - tmp6;
344e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    tmp4 = tmp10 + tmp5;
345e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
346e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    /* Final output stage: scale down by a factor of 8 and range-limit */
347e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
348e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[0] = range_limit[IDESCALE(tmp0 + tmp7, PASS1_BITS+3)
349e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
350e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[7] = range_limit[IDESCALE(tmp0 - tmp7, PASS1_BITS+3)
351e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
352e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[1] = range_limit[IDESCALE(tmp1 + tmp6, PASS1_BITS+3)
353e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
354e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[6] = range_limit[IDESCALE(tmp1 - tmp6, PASS1_BITS+3)
355e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
356e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[2] = range_limit[IDESCALE(tmp2 + tmp5, PASS1_BITS+3)
357e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
358e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[5] = range_limit[IDESCALE(tmp2 - tmp5, PASS1_BITS+3)
359e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
360e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[4] = range_limit[IDESCALE(tmp3 + tmp4, PASS1_BITS+3)
361e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
362e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    outptr[3] = range_limit[IDESCALE(tmp3 - tmp4, PASS1_BITS+3)
363e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov			    & RANGE_MASK];
364e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
365e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov    wsptr += DCTSIZE;		/* advance pointer to next row */
366e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov  }
367e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov}
368e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
369e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif /* DCT_IFAST_SUPPORTED */
370e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov
371e6986e1e8d4a57987f47c215490cb080a65ee29aSvet Ganov#endif //_FX_JPEG_TURBO_
372