nuclear@14: /* nuclear@14: * jfdctfst.c nuclear@14: * nuclear@14: * Copyright (C) 1994-1996, Thomas G. Lane. nuclear@14: * This file is part of the Independent JPEG Group's software. nuclear@14: * For conditions of distribution and use, see the accompanying README file. nuclear@14: * nuclear@14: * This file contains a fast, not so accurate integer implementation of the nuclear@14: * forward DCT (Discrete Cosine Transform). nuclear@14: * nuclear@14: * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT nuclear@14: * on each column. Direct algorithms are also available, but they are nuclear@14: * much more complex and seem not to be any faster when reduced to code. nuclear@14: * nuclear@14: * This implementation is based on Arai, Agui, and Nakajima's algorithm for nuclear@14: * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in nuclear@14: * Japanese, but the algorithm is described in the Pennebaker & Mitchell nuclear@14: * JPEG textbook (see REFERENCES section in file README). The following code nuclear@14: * is based directly on figure 4-8 in P&M. nuclear@14: * While an 8-point DCT cannot be done in less than 11 multiplies, it is nuclear@14: * possible to arrange the computation so that many of the multiplies are nuclear@14: * simple scalings of the final outputs. These multiplies can then be nuclear@14: * folded into the multiplications or divisions by the JPEG quantization nuclear@14: * table entries. The AA&N method leaves only 5 multiplies and 29 adds nuclear@14: * to be done in the DCT itself. nuclear@14: * The primary disadvantage of this method is that with fixed-point math, nuclear@14: * accuracy is lost due to imprecise representation of the scaled nuclear@14: * quantization values. The smaller the quantization table entry, the less nuclear@14: * precise the scaled value, so this implementation does worse with high- nuclear@14: * quality-setting files than with low-quality ones. nuclear@14: */ nuclear@14: nuclear@14: #define JPEG_INTERNALS nuclear@14: #include "jinclude.h" nuclear@14: #include "jpeglib.h" nuclear@14: #include "jdct.h" /* Private declarations for DCT subsystem */ nuclear@14: nuclear@14: #ifdef DCT_IFAST_SUPPORTED nuclear@14: nuclear@14: nuclear@14: /* nuclear@14: * This module is specialized to the case DCTSIZE = 8. nuclear@14: */ nuclear@14: nuclear@14: #if DCTSIZE != 8 nuclear@14: Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ nuclear@14: #endif nuclear@14: nuclear@14: nuclear@14: /* Scaling decisions are generally the same as in the LL&M algorithm; nuclear@14: * see jfdctint.c for more details. However, we choose to descale nuclear@14: * (right shift) multiplication products as soon as they are formed, nuclear@14: * rather than carrying additional fractional bits into subsequent additions. nuclear@14: * This compromises accuracy slightly, but it lets us save a few shifts. nuclear@14: * More importantly, 16-bit arithmetic is then adequate (for 8-bit samples) nuclear@14: * everywhere except in the multiplications proper; this saves a good deal nuclear@14: * of work on 16-bit-int machines. nuclear@14: * nuclear@14: * Again to save a few shifts, the intermediate results between pass 1 and nuclear@14: * pass 2 are not upscaled, but are represented only to integral precision. nuclear@14: * nuclear@14: * A final compromise is to represent the multiplicative constants to only nuclear@14: * 8 fractional bits, rather than 13. This saves some shifting work on some nuclear@14: * machines, and may also reduce the cost of multiplication (since there nuclear@14: * are fewer one-bits in the constants). nuclear@14: */ nuclear@14: nuclear@14: #define CONST_BITS 8 nuclear@14: nuclear@14: nuclear@14: /* Some C compilers fail to reduce "FIX(constant)" at compile time, thus nuclear@14: * causing a lot of useless floating-point operations at run time. nuclear@14: * To get around this we use the following pre-calculated constants. nuclear@14: * If you change CONST_BITS you may want to add appropriate values. nuclear@14: * (With a reasonable C compiler, you can just rely on the FIX() macro...) nuclear@14: */ nuclear@14: nuclear@14: #if CONST_BITS == 8 nuclear@14: #define FIX_0_382683433 ((INT32) 98) /* FIX(0.382683433) */ nuclear@14: #define FIX_0_541196100 ((INT32) 139) /* FIX(0.541196100) */ nuclear@14: #define FIX_0_707106781 ((INT32) 181) /* FIX(0.707106781) */ nuclear@14: #define FIX_1_306562965 ((INT32) 334) /* FIX(1.306562965) */ nuclear@14: #else nuclear@14: #define FIX_0_382683433 FIX(0.382683433) nuclear@14: #define FIX_0_541196100 FIX(0.541196100) nuclear@14: #define FIX_0_707106781 FIX(0.707106781) nuclear@14: #define FIX_1_306562965 FIX(1.306562965) nuclear@14: #endif nuclear@14: nuclear@14: nuclear@14: /* We can gain a little more speed, with a further compromise in accuracy, nuclear@14: * by omitting the addition in a descaling shift. This yields an incorrectly nuclear@14: * rounded result half the time... nuclear@14: */ nuclear@14: nuclear@14: #ifndef USE_ACCURATE_ROUNDING nuclear@14: #undef DESCALE nuclear@14: #define DESCALE(x,n) RIGHT_SHIFT(x, n) nuclear@14: #endif nuclear@14: nuclear@14: nuclear@14: /* Multiply a DCTELEM variable by an INT32 constant, and immediately nuclear@14: * descale to yield a DCTELEM result. nuclear@14: */ nuclear@14: nuclear@14: #define MULTIPLY(var,const) ((DCTELEM) DESCALE((var) * (const), CONST_BITS)) nuclear@14: nuclear@14: nuclear@14: /* nuclear@14: * Perform the forward DCT on one block of samples. nuclear@14: */ nuclear@14: nuclear@14: GLOBAL(void) nuclear@14: jpeg_fdct_ifast (DCTELEM * data) nuclear@14: { nuclear@14: DCTELEM tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; nuclear@14: DCTELEM tmp10, tmp11, tmp12, tmp13; nuclear@14: DCTELEM z1, z2, z3, z4, z5, z11, z13; nuclear@14: DCTELEM *dataptr; nuclear@14: int ctr; nuclear@14: SHIFT_TEMPS nuclear@14: nuclear@14: /* Pass 1: process rows. */ nuclear@14: nuclear@14: dataptr = data; nuclear@14: for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { nuclear@14: tmp0 = dataptr[0] + dataptr[7]; nuclear@14: tmp7 = dataptr[0] - dataptr[7]; nuclear@14: tmp1 = dataptr[1] + dataptr[6]; nuclear@14: tmp6 = dataptr[1] - dataptr[6]; nuclear@14: tmp2 = dataptr[2] + dataptr[5]; nuclear@14: tmp5 = dataptr[2] - dataptr[5]; nuclear@14: tmp3 = dataptr[3] + dataptr[4]; nuclear@14: tmp4 = dataptr[3] - dataptr[4]; nuclear@14: nuclear@14: /* Even part */ nuclear@14: nuclear@14: tmp10 = tmp0 + tmp3; /* phase 2 */ nuclear@14: tmp13 = tmp0 - tmp3; nuclear@14: tmp11 = tmp1 + tmp2; nuclear@14: tmp12 = tmp1 - tmp2; nuclear@14: nuclear@14: dataptr[0] = tmp10 + tmp11; /* phase 3 */ nuclear@14: dataptr[4] = tmp10 - tmp11; nuclear@14: nuclear@14: z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ nuclear@14: dataptr[2] = tmp13 + z1; /* phase 5 */ nuclear@14: dataptr[6] = tmp13 - z1; nuclear@14: nuclear@14: /* Odd part */ nuclear@14: nuclear@14: tmp10 = tmp4 + tmp5; /* phase 2 */ nuclear@14: tmp11 = tmp5 + tmp6; nuclear@14: tmp12 = tmp6 + tmp7; nuclear@14: nuclear@14: /* The rotator is modified from fig 4-8 to avoid extra negations. */ nuclear@14: z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ nuclear@14: z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ nuclear@14: z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ nuclear@14: z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ nuclear@14: nuclear@14: z11 = tmp7 + z3; /* phase 5 */ nuclear@14: z13 = tmp7 - z3; nuclear@14: nuclear@14: dataptr[5] = z13 + z2; /* phase 6 */ nuclear@14: dataptr[3] = z13 - z2; nuclear@14: dataptr[1] = z11 + z4; nuclear@14: dataptr[7] = z11 - z4; nuclear@14: nuclear@14: dataptr += DCTSIZE; /* advance pointer to next row */ nuclear@14: } nuclear@14: nuclear@14: /* Pass 2: process columns. */ nuclear@14: nuclear@14: dataptr = data; nuclear@14: for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { nuclear@14: tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; nuclear@14: tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; nuclear@14: tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; nuclear@14: tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; nuclear@14: tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; nuclear@14: tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; nuclear@14: tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; nuclear@14: tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; nuclear@14: nuclear@14: /* Even part */ nuclear@14: nuclear@14: tmp10 = tmp0 + tmp3; /* phase 2 */ nuclear@14: tmp13 = tmp0 - tmp3; nuclear@14: tmp11 = tmp1 + tmp2; nuclear@14: tmp12 = tmp1 - tmp2; nuclear@14: nuclear@14: dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */ nuclear@14: dataptr[DCTSIZE*4] = tmp10 - tmp11; nuclear@14: nuclear@14: z1 = MULTIPLY(tmp12 + tmp13, FIX_0_707106781); /* c4 */ nuclear@14: dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */ nuclear@14: dataptr[DCTSIZE*6] = tmp13 - z1; nuclear@14: nuclear@14: /* Odd part */ nuclear@14: nuclear@14: tmp10 = tmp4 + tmp5; /* phase 2 */ nuclear@14: tmp11 = tmp5 + tmp6; nuclear@14: tmp12 = tmp6 + tmp7; nuclear@14: nuclear@14: /* The rotator is modified from fig 4-8 to avoid extra negations. */ nuclear@14: z5 = MULTIPLY(tmp10 - tmp12, FIX_0_382683433); /* c6 */ nuclear@14: z2 = MULTIPLY(tmp10, FIX_0_541196100) + z5; /* c2-c6 */ nuclear@14: z4 = MULTIPLY(tmp12, FIX_1_306562965) + z5; /* c2+c6 */ nuclear@14: z3 = MULTIPLY(tmp11, FIX_0_707106781); /* c4 */ nuclear@14: nuclear@14: z11 = tmp7 + z3; /* phase 5 */ nuclear@14: z13 = tmp7 - z3; nuclear@14: nuclear@14: dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */ nuclear@14: dataptr[DCTSIZE*3] = z13 - z2; nuclear@14: dataptr[DCTSIZE*1] = z11 + z4; nuclear@14: dataptr[DCTSIZE*7] = z11 - z4; nuclear@14: nuclear@14: dataptr++; /* advance pointer to next column */ nuclear@14: } nuclear@14: } nuclear@14: nuclear@14: #endif /* DCT_IFAST_SUPPORTED */