yading@10: /* yading@10: * This file is part of the Independent JPEG Group's software. yading@10: * yading@10: * The authors make NO WARRANTY or representation, either express or implied, yading@10: * with respect to this software, its quality, accuracy, merchantability, or yading@10: * fitness for a particular purpose. This software is provided "AS IS", and yading@10: * you, its user, assume the entire risk as to its quality and accuracy. yading@10: * yading@10: * This software is copyright (C) 1994-1996, Thomas G. Lane. yading@10: * All Rights Reserved except as specified below. yading@10: * yading@10: * Permission is hereby granted to use, copy, modify, and distribute this yading@10: * software (or portions thereof) for any purpose, without fee, subject to yading@10: * these conditions: yading@10: * (1) If any part of the source code for this software is distributed, then yading@10: * this README file must be included, with this copyright and no-warranty yading@10: * notice unaltered; and any additions, deletions, or changes to the original yading@10: * files must be clearly indicated in accompanying documentation. yading@10: * (2) If only executable code is distributed, then the accompanying yading@10: * documentation must state that "this software is based in part on the work yading@10: * of the Independent JPEG Group". yading@10: * (3) Permission for use of this software is granted only if the user accepts yading@10: * full responsibility for any undesirable consequences; the authors accept yading@10: * NO LIABILITY for damages of any kind. yading@10: * yading@10: * These conditions apply to any software derived from or based on the IJG yading@10: * code, not just to the unmodified library. If you use our work, you ought yading@10: * to acknowledge us. yading@10: * yading@10: * Permission is NOT granted for the use of any IJG author's name or company yading@10: * name in advertising or publicity relating to this software or products yading@10: * derived from it. This software may be referred to only as "the Independent yading@10: * JPEG Group's software". yading@10: * yading@10: * We specifically permit and encourage the use of this software as the basis yading@10: * of commercial products, provided that all warranty or liability claims are yading@10: * assumed by the product vendor. yading@10: * yading@10: * This file contains a fast, not so accurate integer implementation of the yading@10: * forward DCT (Discrete Cosine Transform). yading@10: * yading@10: * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT yading@10: * on each column. Direct algorithms are also available, but they are yading@10: * much more complex and seem not to be any faster when reduced to code. yading@10: * yading@10: * This implementation is based on Arai, Agui, and Nakajima's algorithm for yading@10: * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in yading@10: * Japanese, but the algorithm is described in the Pennebaker & Mitchell yading@10: * JPEG textbook (see REFERENCES section in file README). The following code yading@10: * is based directly on figure 4-8 in P&M. yading@10: * While an 8-point DCT cannot be done in less than 11 multiplies, it is yading@10: * possible to arrange the computation so that many of the multiplies are yading@10: * simple scalings of the final outputs. These multiplies can then be yading@10: * folded into the multiplications or divisions by the JPEG quantization yading@10: * table entries. The AA&N method leaves only 5 multiplies and 29 adds yading@10: * to be done in the DCT itself. yading@10: * The primary disadvantage of this method is that with fixed-point math, yading@10: * accuracy is lost due to imprecise representation of the scaled yading@10: * quantization values. The smaller the quantization table entry, the less yading@10: * precise the scaled value, so this implementation does worse with high- yading@10: * quality-setting files than with low-quality ones. yading@10: */ yading@10: yading@10: /** yading@10: * @file yading@10: * Independent JPEG Group's fast AAN dct. yading@10: */ yading@10: yading@10: #include yading@10: #include yading@10: #include "libavutil/common.h" yading@10: #include "dct.h" yading@10: yading@10: #define DCTSIZE 8 yading@10: #define GLOBAL(x) x yading@10: #define RIGHT_SHIFT(x, n) ((x) >> (n)) yading@10: yading@10: /* yading@10: * This module is specialized to the case DCTSIZE = 8. yading@10: */ yading@10: yading@10: #if DCTSIZE != 8 yading@10: Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ yading@10: #endif yading@10: yading@10: yading@10: /* Scaling decisions are generally the same as in the LL&M algorithm; yading@10: * see jfdctint.c for more details. However, we choose to descale yading@10: * (right shift) multiplication products as soon as they are formed, yading@10: * rather than carrying additional fractional bits into subsequent additions. yading@10: * This compromises accuracy slightly, but it lets us save a few shifts. yading@10: * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples) yading@10: * everywhere except in the multiplications proper; this saves a good deal yading@10: * of work on 16-bit-int machines. yading@10: * yading@10: * Again to save a few shifts, the intermediate results between pass 1 and yading@10: * pass 2 are not upscaled, but are represented only to integral precision. yading@10: * yading@10: * A final compromise is to represent the multiplicative constants to only yading@10: * 8 fractional bits, rather than 13. This saves some shifting work on some yading@10: * machines, and may also reduce the cost of multiplication (since there yading@10: * are fewer one-bits in the constants). yading@10: */ yading@10: yading@10: #define CONST_BITS 8 yading@10: yading@10: yading@10: /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus yading@10: * causing a lot of useless floating-point operations at run time. yading@10: * To get around this we use the following pre-calculated constants. yading@10: * If you change CONST_BITS you may want to add appropriate values. yading@10: * (With a reasonable C compiler, you can just rely on the FIX() macro...) yading@10: */ yading@10: yading@10: #if CONST_BITS == 8 yading@10: #define FIX_0_382683433 ((int32_t) 98) /* FIX(0.382683433) */ yading@10: #define FIX_0_541196100 ((int32_t) 139) /* FIX(0.541196100) */ yading@10: #define FIX_0_707106781 ((int32_t) 181) /* FIX(0.707106781) */ yading@10: #define FIX_1_306562965 ((int32_t) 334) /* FIX(1.306562965) */ yading@10: #else yading@10: #define FIX_0_382683433 FIX(0.382683433) yading@10: #define FIX_0_541196100 FIX(0.541196100) yading@10: #define FIX_0_707106781 FIX(0.707106781) yading@10: #define FIX_1_306562965 FIX(1.306562965) yading@10: #endif yading@10: yading@10: yading@10: /* We can gain a little more speed, with a further compromise in accuracy, yading@10: * by omitting the addition in a descaling shift. This yields an incorrectly yading@10: * rounded result half the time... yading@10: */ yading@10: yading@10: #ifndef USE_ACCURATE_ROUNDING yading@10: #undef DESCALE yading@10: #define DESCALE(x,n) RIGHT_SHIFT(x, n) yading@10: #endif yading@10: yading@10: yading@10: /* Multiply a int16_t variable by an int32_t constant, and immediately yading@10: * descale to yield a int16_t result. yading@10: */ yading@10: yading@10: #define MULTIPLY(var,const) ((int16_t) DESCALE((var) * (const), CONST_BITS)) yading@10: yading@10: static av_always_inline void row_fdct(int16_t * data){ yading@10: int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; yading@10: int tmp10, tmp11, tmp12, tmp13; yading@10: int z1, z2, z3, z4, z5, z11, z13; yading@10: int16_t *dataptr; yading@10: int ctr; yading@10: yading@10: /* Pass 1: process rows. */ yading@10: yading@10: dataptr = data; yading@10: for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { yading@10: tmp0 = dataptr[0] + dataptr[7]; yading@10: tmp7 = dataptr[0] - dataptr[7]; yading@10: tmp1 = dataptr[1] + dataptr[6]; yading@10: tmp6 = dataptr[1] - dataptr[6]; yading@10: tmp2 = dataptr[2] + dataptr[5]; yading@10: tmp5 = dataptr[2] - dataptr[5]; yading@10: tmp3 = dataptr[3] + dataptr[4]; yading@10: tmp4 = dataptr[3] - dataptr[4]; yading@10: yading@10: /* Even part */ yading@10: yading@10: tmp10 = tmp0 + tmp3; /* phase 2 */ yading@10: tmp13 = tmp0 - tmp3; yading@10: tmp11 = tmp1 + tmp2; yading@10: tmp12 = tmp1 - tmp2; yading@10: yading@10: dataptr[0] = tmp10 + tmp11; /* phase 3 */ yading@10: dataptr[4] = tmp10 - tmp11; yading@10: yading@10: z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ yading@10: dataptr[2] = tmp13 + z1; /* phase 5 */ yading@10: dataptr[6] = tmp13 - z1; yading@10: yading@10: /* Odd part */ yading@10: yading@10: tmp10 = tmp4 + tmp5; /* phase 2 */ yading@10: tmp11 = tmp5 + tmp6; yading@10: tmp12 = tmp6 + tmp7; yading@10: yading@10: /* The rotator is modified from fig 4-8 to avoid extra negations. */ yading@10: z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ yading@10: z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ yading@10: z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ yading@10: z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ yading@10: yading@10: z11 = tmp7 + z3; /* phase 5 */ yading@10: z13 = tmp7 - z3; yading@10: yading@10: dataptr[5] = z13 + z2; /* phase 6 */ yading@10: dataptr[3] = z13 - z2; yading@10: dataptr[1] = z11 + z4; yading@10: dataptr[7] = z11 - z4; yading@10: yading@10: dataptr += DCTSIZE; /* advance pointer to next row */ yading@10: } yading@10: } yading@10: yading@10: /* yading@10: * Perform the forward DCT on one block of samples. yading@10: */ yading@10: yading@10: GLOBAL(void) yading@10: ff_fdct_ifast (int16_t * data) yading@10: { yading@10: int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; yading@10: int tmp10, tmp11, tmp12, tmp13; yading@10: int z1, z2, z3, z4, z5, z11, z13; yading@10: int16_t *dataptr; yading@10: int ctr; yading@10: yading@10: row_fdct(data); yading@10: yading@10: /* Pass 2: process columns. */ yading@10: yading@10: dataptr = data; yading@10: for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { yading@10: tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; yading@10: tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; yading@10: tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; yading@10: tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; yading@10: tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; yading@10: tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; yading@10: tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; yading@10: tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; yading@10: yading@10: /* Even part */ yading@10: yading@10: tmp10 = tmp0 + tmp3; /* phase 2 */ yading@10: tmp13 = tmp0 - tmp3; yading@10: tmp11 = tmp1 + tmp2; yading@10: tmp12 = tmp1 - tmp2; yading@10: yading@10: dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */ yading@10: dataptr[DCTSIZE*4] = tmp10 - tmp11; yading@10: yading@10: z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ yading@10: dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */ yading@10: dataptr[DCTSIZE*6] = tmp13 - z1; yading@10: yading@10: /* Odd part */ yading@10: yading@10: tmp10 = tmp4 + tmp5; /* phase 2 */ yading@10: tmp11 = tmp5 + tmp6; yading@10: tmp12 = tmp6 + tmp7; yading@10: yading@10: /* The rotator is modified from fig 4-8 to avoid extra negations. */ yading@10: z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ yading@10: z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ yading@10: z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ yading@10: z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ yading@10: yading@10: z11 = tmp7 + z3; /* phase 5 */ yading@10: z13 = tmp7 - z3; yading@10: yading@10: dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */ yading@10: dataptr[DCTSIZE*3] = z13 - z2; yading@10: dataptr[DCTSIZE*1] = z11 + z4; yading@10: dataptr[DCTSIZE*7] = z11 - z4; yading@10: yading@10: dataptr++; /* advance pointer to next column */ yading@10: } yading@10: } yading@10: yading@10: /* yading@10: * Perform the forward 2-4-8 DCT on one block of samples. yading@10: */ yading@10: yading@10: GLOBAL(void) yading@10: ff_fdct_ifast248 (int16_t * data) yading@10: { yading@10: int tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; yading@10: int tmp10, tmp11, tmp12, tmp13; yading@10: int z1; yading@10: int16_t *dataptr; yading@10: int ctr; yading@10: yading@10: row_fdct(data); yading@10: yading@10: /* Pass 2: process columns. */ yading@10: yading@10: dataptr = data; yading@10: for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { yading@10: tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*1]; yading@10: tmp1 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*3]; yading@10: tmp2 = dataptr[DCTSIZE*4] + dataptr[DCTSIZE*5]; yading@10: tmp3 = dataptr[DCTSIZE*6] + dataptr[DCTSIZE*7]; yading@10: tmp4 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*1]; yading@10: tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*3]; yading@10: tmp6 = dataptr[DCTSIZE*4] - dataptr[DCTSIZE*5]; yading@10: tmp7 = dataptr[DCTSIZE*6] - dataptr[DCTSIZE*7]; yading@10: yading@10: /* Even part */ yading@10: yading@10: tmp10 = tmp0 + tmp3; yading@10: tmp11 = tmp1 + tmp2; yading@10: tmp12 = tmp1 - tmp2; yading@10: tmp13 = tmp0 - tmp3; yading@10: yading@10: dataptr[DCTSIZE*0] = tmp10 + tmp11; yading@10: dataptr[DCTSIZE*4] = tmp10 - tmp11; yading@10: yading@10: z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); yading@10: dataptr[DCTSIZE*2] = tmp13 + z1; yading@10: dataptr[DCTSIZE*6] = tmp13 - z1; yading@10: yading@10: tmp10 = tmp4 + tmp7; yading@10: tmp11 = tmp5 + tmp6; yading@10: tmp12 = tmp5 - tmp6; yading@10: tmp13 = tmp4 - tmp7; yading@10: yading@10: dataptr[DCTSIZE*1] = tmp10 + tmp11; yading@10: dataptr[DCTSIZE*5] = tmp10 - tmp11; yading@10: yading@10: z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); yading@10: dataptr[DCTSIZE*3] = tmp13 + z1; yading@10: dataptr[DCTSIZE*7] = tmp13 - z1; yading@10: yading@10: dataptr++; /* advance pointer to next column */ yading@10: } yading@10: } yading@10: yading@10: yading@10: #undef GLOBAL yading@10: #undef CONST_BITS yading@10: #undef DESCALE yading@10: #undef FIX_0_541196100 yading@10: #undef FIX_1_306562965