Raghu Gandham | 71ee859 | 2012-11-06 10:17:40 -0800 | [diff] [blame] | 1 | /* |
| 2 | * IDCT implementation using the MIPS DSP ASE (little endian version) |
| 3 | * |
| 4 | * jidctfst.c |
| 5 | * |
| 6 | * Copyright (C) 1994-1998, Thomas G. Lane. |
| 7 | * This file is part of the Independent JPEG Group's software. |
| 8 | * For conditions of distribution and use, see the accompanying README file. |
| 9 | * |
| 10 | * This file contains a fast, not so accurate integer implementation of the |
| 11 | * inverse DCT (Discrete Cosine Transform). In the IJG code, this routine |
| 12 | * must also perform dequantization of the input coefficients. |
| 13 | * |
| 14 | * A 2-D IDCT can be done by 1-D IDCT on each column followed by 1-D IDCT |
| 15 | * on each row (or vice versa, but it's more convenient to emit a row at |
| 16 | * a time). Direct algorithms are also available, but they are much more |
| 17 | * complex and seem not to be any faster when reduced to code. |
| 18 | * |
| 19 | * This implementation is based on Arai, Agui, and Nakajima's algorithm for |
| 20 | * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in |
| 21 | * Japanese, but the algorithm is described in the Pennebaker & Mitchell |
| 22 | * JPEG textbook (see REFERENCES section in file README). The following code |
| 23 | * is based directly on figure 4-8 in P&M. |
| 24 | * While an 8-point DCT cannot be done in less than 11 multiplies, it is |
| 25 | * possible to arrange the computation so that many of the multiplies are |
| 26 | * simple scalings of the final outputs. These multiplies can then be |
| 27 | * folded into the multiplications or divisions by the JPEG quantization |
| 28 | * table entries. The AA&N method leaves only 5 multiplies and 29 adds |
| 29 | * to be done in the DCT itself. |
| 30 | * The primary disadvantage of this method is that with fixed-point math, |
| 31 | * accuracy is lost due to imprecise representation of the scaled |
| 32 | * quantization values. The smaller the quantization table entry, the less |
| 33 | * precise the scaled value, so this implementation does worse with high- |
| 34 | * quality-setting files than with low-quality ones. |
| 35 | */ |
| 36 | |
| 37 | #define JPEG_INTERNALS |
| 38 | #include "jinclude.h" |
| 39 | #include "jpeglib.h" |
| 40 | #include "jdct.h" /* Private declarations for DCT subsystem */ |
| 41 | |
| 42 | #ifdef DCT_IFAST_SUPPORTED |
| 43 | |
| 44 | |
| 45 | /* |
| 46 | * This module is specialized to the case DCTSIZE = 8. |
| 47 | */ |
| 48 | |
| 49 | #if DCTSIZE != 8 |
| 50 | Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ |
| 51 | #endif |
| 52 | |
| 53 | |
| 54 | /* Scaling decisions are generally the same as in the LL&M algorithm; |
| 55 | * see jidctint.c for more details. However, we choose to descale |
| 56 | * (right shift) multiplication products as soon as they are formed, |
| 57 | * rather than carrying additional fractional bits into subsequent additions. |
| 58 | * This compromises accuracy slightly, but it lets us save a few shifts. |
| 59 | * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples) |
| 60 | * everywhere except in the multiplications proper; this saves a good deal |
| 61 | * of work on 16-bit-int machines. |
| 62 | * |
| 63 | * The dequantized coefficients are not integers because the AA&N scaling |
| 64 | * factors have been incorporated. We represent them scaled up by PASS1_BITS, |
| 65 | * so that the first and second IDCT rounds have the same input scaling. |
| 66 | * For 8-bit JSAMPLEs, we choose IFAST_SCALE_BITS = PASS1_BITS so as to |
| 67 | * avoid a descaling shift; this compromises accuracy rather drastically |
| 68 | * for small quantization table entries, but it saves a lot of shifts. |
| 69 | * For 12-bit JSAMPLEs, there's no hope of using 16x16 multiplies anyway, |
| 70 | * so we use a much larger scaling factor to preserve accuracy. |
| 71 | * |
| 72 | * A final compromise is to represent the multiplicative constants to only |
| 73 | * 8 fractional bits, rather than 13. This saves some shifting work on some |
| 74 | * machines, and may also reduce the cost of multiplication (since there |
| 75 | * are fewer one-bits in the constants). |
| 76 | */ |
| 77 | |
| 78 | #if BITS_IN_JSAMPLE == 8 |
| 79 | #define CONST_BITS 8 |
| 80 | #define PASS1_BITS 2 |
| 81 | #else |
| 82 | #define CONST_BITS 8 |
| 83 | #define PASS1_BITS 1 /* lose a little precision to avoid overflow */ |
| 84 | #endif |
| 85 | |
| 86 | /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus |
| 87 | * causing a lot of useless floating-point operations at run time. |
| 88 | * To get around this we use the following pre-calculated constants. |
| 89 | * If you change CONST_BITS you may want to add appropriate values. |
| 90 | * (With a reasonable C compiler, you can just rely on the FIX() macro...) |
| 91 | */ |
| 92 | |
| 93 | #if CONST_BITS == 8 |
| 94 | #define FIX_1_082392200 ((INT32) 277) /* FIX(1.082392200) */ |
| 95 | #define FIX_1_414213562 ((INT32) 362) /* FIX(1.414213562) */ |
| 96 | #define FIX_1_847759065 ((INT32) 473) /* FIX(1.847759065) */ |
| 97 | #define FIX_2_613125930 ((INT32) 669) /* FIX(2.613125930) */ |
| 98 | #else |
| 99 | #define FIX_1_082392200 FIX(1.082392200) |
| 100 | #define FIX_1_414213562 FIX(1.414213562) |
| 101 | #define FIX_1_847759065 FIX(1.847759065) |
| 102 | #define FIX_2_613125930 FIX(2.613125930) |
| 103 | #endif |
| 104 | |
| 105 | |
| 106 | /* We can gain a little more speed, with a further compromise in accuracy, |
| 107 | * by omitting the addition in a descaling shift. This yields an incorrectly |
| 108 | * rounded result half the time... |
| 109 | */ |
| 110 | |
| 111 | #ifndef USE_ACCURATE_ROUNDING |
| 112 | #undef DESCALE |
| 113 | #define DESCALE(x,n) RIGHT_SHIFT(x, n) |
| 114 | #endif |
| 115 | |
| 116 | |
| 117 | /* Multiply a DCTELEM variable by an INT32 constant, and immediately |
| 118 | * descale to yield a DCTELEM result. |
| 119 | */ |
| 120 | |
| 121 | #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS)) |
| 122 | |
| 123 | |
| 124 | /* Dequantize a coefficient by multiplying it by the multiplier-table |
| 125 | * entry; produce a DCTELEM result. For 8-bit data a 16x16->16 |
| 126 | * multiplication will do. For 12-bit data, the multiplier table is |
| 127 | * declared INT32, so a 32-bit multiply will be used. |
| 128 | */ |
| 129 | |
| 130 | #if BITS_IN_JSAMPLE == 8 |
| 131 | #define DEQUANTIZE(coef,quantval) (((IFAST_MULT_TYPE) (coef)) * (quantval)) |
| 132 | #else |
| 133 | #define DEQUANTIZE(coef,quantval) \ |
| 134 | DESCALE((coef)*(quantval), IFAST_SCALE_BITS-PASS1_BITS) |
| 135 | #endif |
| 136 | |
| 137 | |
| 138 | /* Like DESCALE, but applies to a DCTELEM and produces an int. |
| 139 | * We assume that int right shift is unsigned if INT32 right shift is. |
| 140 | */ |
| 141 | |
| 142 | #ifdef RIGHT_SHIFT_IS_UNSIGNED |
| 143 | #define ISHIFT_TEMPS DCTELEM ishift_temp; |
| 144 | #if BITS_IN_JSAMPLE == 8 |
| 145 | #define DCTELEMBITS 16 /* DCTELEM may be 16 or 32 bits */ |
| 146 | #else |
| 147 | #define DCTELEMBITS 32 /* DCTELEM must be 32 bits */ |
| 148 | #endif |
| 149 | #define IRIGHT_SHIFT(x,shft) \ |
| 150 | ((ishift_temp = (x)) < 0 ? \ |
| 151 | (ishift_temp >> (shft)) | ((~((DCTELEM) 0)) << (DCTELEMBITS-(shft))) : \ |
| 152 | (ishift_temp >> (shft))) |
| 153 | #else |
| 154 | #define ISHIFT_TEMPS |
| 155 | #define IRIGHT_SHIFT(x,shft) ((x) >> (shft)) |
| 156 | #endif |
| 157 | |
| 158 | #ifdef USE_ACCURATE_ROUNDING |
| 159 | #define IDESCALE(x,n) ((int) IRIGHT_SHIFT((x) + (1 << ((n)-1)), n)) |
| 160 | #else |
| 161 | #define IDESCALE(x,n) ((int) IRIGHT_SHIFT(x, n)) |
| 162 | #endif |
| 163 | |
| 164 | |
| 165 | // this table of constants has been moved from mips_idct_le/_be.s to |
| 166 | // avoid having to make the assembler code position independent |
| 167 | static const int mips_idct_coefs[4] = { |
| 168 | 0x45464546, // FIX( 1.082392200 / 2) = 17734 = 0x4546 |
| 169 | 0x5A825A82, // FIX( 1.414213562 / 2) = 23170 = 0x5A82 |
| 170 | 0x76427642, // FIX( 1.847759065 / 2) = 30274 = 0x7642 |
| 171 | 0xAC61AC61 // FIX(-2.613125930 / 4) = -21407 = 0xAC61 |
| 172 | }; |
| 173 | |
| 174 | void mips_idct_columns(JCOEF * inptr, IFAST_MULT_TYPE * quantptr, |
| 175 | DCTELEM * wsptr, const int * mips_idct_coefs); |
| 176 | void mips_idct_rows(DCTELEM * wsptr, JSAMPARRAY output_buf, |
| 177 | JDIMENSION output_col, const int * mips_idct_coefs); |
| 178 | |
| 179 | |
| 180 | /* |
| 181 | * Perform dequantization and inverse DCT on one block of coefficients. |
| 182 | */ |
| 183 | |
| 184 | GLOBAL(void) |
| 185 | jpeg_idct_mips (j_decompress_ptr cinfo, jpeg_component_info * compptr, |
| 186 | JCOEFPTR coef_block, |
| 187 | JSAMPARRAY output_buf, JDIMENSION output_col) |
| 188 | { |
| 189 | JCOEFPTR inptr; |
| 190 | IFAST_MULT_TYPE * quantptr; |
| 191 | DCTELEM workspace[DCTSIZE2]; /* buffers data between passes */ |
| 192 | |
| 193 | /* Pass 1: process columns from input, store into work array. */ |
| 194 | |
| 195 | inptr = coef_block; |
| 196 | quantptr = (IFAST_MULT_TYPE *) compptr->dct_table; |
| 197 | |
| 198 | mips_idct_columns(inptr, quantptr, workspace, mips_idct_coefs); |
| 199 | |
| 200 | /* Pass 2: process rows from work array, store into output array. */ |
| 201 | /* Note that we must descale the results by a factor of 8 == 2**3, */ |
| 202 | /* and also undo the PASS1_BITS scaling. */ |
| 203 | |
| 204 | mips_idct_rows(workspace, output_buf, output_col, mips_idct_coefs); |
| 205 | |
| 206 | } |
| 207 | |
| 208 | #endif /* DCT_IFAST_SUPPORTED */ |