media/libjpeg/jfdctflt.c@b8a032363ba2 (annotated)
media/libjpeg/jfdctflt.c
Thu, 22 Jan 2015 13:21:57 +0100
- author
- Michael Schloh von Bennewitz <michael@schloh.com>
- date
- Thu, 22 Jan 2015 13:21:57 +0100
- branch
- TOR_BUG_9701
- changeset 15
- b8a032363ba2
- permissions
- -rw-r--r--
Incorporate requested changes from Mozilla in review:
https://bugzilla.mozilla.org/show_bug.cgi?id=1123480#c6
| michael@0 | 1 | /* |
| michael@0 | 2 | * jfdctflt.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 floating-point implementation of the |
| michael@0 | 9 | * forward DCT (Discrete Cosine Transform). |
| michael@0 | 10 | * |
| michael@0 | 11 | * This implementation should be more accurate than either of the integer |
| michael@0 | 12 | * DCT implementations. However, it may not give the same results on all |
| michael@0 | 13 | * machines because of differences in roundoff behavior. Speed will depend |
| michael@0 | 14 | * on the hardware's floating point capacity. |
| michael@0 | 15 | * |
| michael@0 | 16 | * A 2-D DCT can be done by 1-D DCT on each row followed by 1-D DCT |
| michael@0 | 17 | * on each column. Direct algorithms are also available, but they are |
| michael@0 | 18 | * much more complex and seem not to be any faster when reduced to code. |
| michael@0 | 19 | * |
| michael@0 | 20 | * This implementation is based on Arai, Agui, and Nakajima's algorithm for |
| michael@0 | 21 | * scaled DCT. Their original paper (Trans. IEICE E-71(11):1095) is in |
| michael@0 | 22 | * Japanese, but the algorithm is described in the Pennebaker & Mitchell |
| michael@0 | 23 | * JPEG textbook (see REFERENCES section in file README). The following code |
| michael@0 | 24 | * is based directly on figure 4-8 in P&M. |
| michael@0 | 25 | * While an 8-point DCT cannot be done in less than 11 multiplies, it is |
| michael@0 | 26 | * possible to arrange the computation so that many of the multiplies are |
| michael@0 | 27 | * simple scalings of the final outputs. These multiplies can then be |
| michael@0 | 28 | * folded into the multiplications or divisions by the JPEG quantization |
| michael@0 | 29 | * table entries. The AA&N method leaves only 5 multiplies and 29 adds |
| michael@0 | 30 | * to be done in the DCT itself. |
| michael@0 | 31 | * The primary disadvantage of this method is that with a fixed-point |
| michael@0 | 32 | * implementation, accuracy is lost due to imprecise representation of the |
| michael@0 | 33 | * scaled quantization values. However, that problem does not arise if |
| michael@0 | 34 | * we use floating point arithmetic. |
| michael@0 | 35 | */ |
| michael@0 | 36 | |
| michael@0 | 37 | #define JPEG_INTERNALS |
| michael@0 | 38 | #include "jinclude.h" |
| michael@0 | 39 | #include "jpeglib.h" |
| michael@0 | 40 | #include "jdct.h" /* Private declarations for DCT subsystem */ |
| michael@0 | 41 | |
| michael@0 | 42 | #ifdef DCT_FLOAT_SUPPORTED |
| michael@0 | 43 | |
| michael@0 | 44 | |
| michael@0 | 45 | /* |
| michael@0 | 46 | * This module is specialized to the case DCTSIZE = 8. |
| michael@0 | 47 | */ |
| michael@0 | 48 | |
| michael@0 | 49 | #if DCTSIZE != 8 |
| michael@0 | 50 | Sorry, this code only copes with 8x8 DCTs. /* deliberate syntax err */ |
| michael@0 | 51 | #endif |
| michael@0 | 52 | |
| michael@0 | 53 | |
| michael@0 | 54 | /* |
| michael@0 | 55 | * Perform the forward DCT on one block of samples. |
| michael@0 | 56 | */ |
| michael@0 | 57 | |
| michael@0 | 58 | GLOBAL(void) |
| michael@0 | 59 | jpeg_fdct_float (FAST_FLOAT * data) |
| michael@0 | 60 | { |
| michael@0 | 61 | FAST_FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7; |
| michael@0 | 62 | FAST_FLOAT tmp10, tmp11, tmp12, tmp13; |
| michael@0 | 63 | FAST_FLOAT z1, z2, z3, z4, z5, z11, z13; |
| michael@0 | 64 | FAST_FLOAT *dataptr; |
| michael@0 | 65 | int ctr; |
| michael@0 | 66 | |
| michael@0 | 67 | /* Pass 1: process rows. */ |
| michael@0 | 68 | |
| michael@0 | 69 | dataptr = data; |
| michael@0 | 70 | for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { |
| michael@0 | 71 | tmp0 = dataptr[0] + dataptr[7]; |
| michael@0 | 72 | tmp7 = dataptr[0] - dataptr[7]; |
| michael@0 | 73 | tmp1 = dataptr[1] + dataptr[6]; |
| michael@0 | 74 | tmp6 = dataptr[1] - dataptr[6]; |
| michael@0 | 75 | tmp2 = dataptr[2] + dataptr[5]; |
| michael@0 | 76 | tmp5 = dataptr[2] - dataptr[5]; |
| michael@0 | 77 | tmp3 = dataptr[3] + dataptr[4]; |
| michael@0 | 78 | tmp4 = dataptr[3] - dataptr[4]; |
| michael@0 | 79 | |
| michael@0 | 80 | /* Even part */ |
| michael@0 | 81 | |
| michael@0 | 82 | tmp10 = tmp0 + tmp3; /* phase 2 */ |
| michael@0 | 83 | tmp13 = tmp0 - tmp3; |
| michael@0 | 84 | tmp11 = tmp1 + tmp2; |
| michael@0 | 85 | tmp12 = tmp1 - tmp2; |
| michael@0 | 86 | |
| michael@0 | 87 | dataptr[0] = tmp10 + tmp11; /* phase 3 */ |
| michael@0 | 88 | dataptr[4] = tmp10 - tmp11; |
| michael@0 | 89 | |
| michael@0 | 90 | z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */ |
| michael@0 | 91 | dataptr[2] = tmp13 + z1; /* phase 5 */ |
| michael@0 | 92 | dataptr[6] = tmp13 - z1; |
| michael@0 | 93 | |
| michael@0 | 94 | /* Odd part */ |
| michael@0 | 95 | |
| michael@0 | 96 | tmp10 = tmp4 + tmp5; /* phase 2 */ |
| michael@0 | 97 | tmp11 = tmp5 + tmp6; |
| michael@0 | 98 | tmp12 = tmp6 + tmp7; |
| michael@0 | 99 | |
| michael@0 | 100 | /* The rotator is modified from fig 4-8 to avoid extra negations. */ |
| michael@0 | 101 | z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */ |
| michael@0 | 102 | z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */ |
| michael@0 | 103 | z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */ |
| michael@0 | 104 | z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */ |
| michael@0 | 105 | |
| michael@0 | 106 | z11 = tmp7 + z3; /* phase 5 */ |
| michael@0 | 107 | z13 = tmp7 - z3; |
| michael@0 | 108 | |
| michael@0 | 109 | dataptr[5] = z13 + z2; /* phase 6 */ |
| michael@0 | 110 | dataptr[3] = z13 - z2; |
| michael@0 | 111 | dataptr[1] = z11 + z4; |
| michael@0 | 112 | dataptr[7] = z11 - z4; |
| michael@0 | 113 | |
| michael@0 | 114 | dataptr += DCTSIZE; /* advance pointer to next row */ |
| michael@0 | 115 | } |
| michael@0 | 116 | |
| michael@0 | 117 | /* Pass 2: process columns. */ |
| michael@0 | 118 | |
| michael@0 | 119 | dataptr = data; |
| michael@0 | 120 | for (ctr = DCTSIZE-1; ctr >= 0; ctr--) { |
| michael@0 | 121 | tmp0 = dataptr[DCTSIZE*0] + dataptr[DCTSIZE*7]; |
| michael@0 | 122 | tmp7 = dataptr[DCTSIZE*0] - dataptr[DCTSIZE*7]; |
| michael@0 | 123 | tmp1 = dataptr[DCTSIZE*1] + dataptr[DCTSIZE*6]; |
| michael@0 | 124 | tmp6 = dataptr[DCTSIZE*1] - dataptr[DCTSIZE*6]; |
| michael@0 | 125 | tmp2 = dataptr[DCTSIZE*2] + dataptr[DCTSIZE*5]; |
| michael@0 | 126 | tmp5 = dataptr[DCTSIZE*2] - dataptr[DCTSIZE*5]; |
| michael@0 | 127 | tmp3 = dataptr[DCTSIZE*3] + dataptr[DCTSIZE*4]; |
| michael@0 | 128 | tmp4 = dataptr[DCTSIZE*3] - dataptr[DCTSIZE*4]; |
| michael@0 | 129 | |
| michael@0 | 130 | /* Even part */ |
| michael@0 | 131 | |
| michael@0 | 132 | tmp10 = tmp0 + tmp3; /* phase 2 */ |
| michael@0 | 133 | tmp13 = tmp0 - tmp3; |
| michael@0 | 134 | tmp11 = tmp1 + tmp2; |
| michael@0 | 135 | tmp12 = tmp1 - tmp2; |
| michael@0 | 136 | |
| michael@0 | 137 | dataptr[DCTSIZE*0] = tmp10 + tmp11; /* phase 3 */ |
| michael@0 | 138 | dataptr[DCTSIZE*4] = tmp10 - tmp11; |
| michael@0 | 139 | |
| michael@0 | 140 | z1 = (tmp12 + tmp13) * ((FAST_FLOAT) 0.707106781); /* c4 */ |
| michael@0 | 141 | dataptr[DCTSIZE*2] = tmp13 + z1; /* phase 5 */ |
| michael@0 | 142 | dataptr[DCTSIZE*6] = tmp13 - z1; |
| michael@0 | 143 | |
| michael@0 | 144 | /* Odd part */ |
| michael@0 | 145 | |
| michael@0 | 146 | tmp10 = tmp4 + tmp5; /* phase 2 */ |
| michael@0 | 147 | tmp11 = tmp5 + tmp6; |
| michael@0 | 148 | tmp12 = tmp6 + tmp7; |
| michael@0 | 149 | |
| michael@0 | 150 | /* The rotator is modified from fig 4-8 to avoid extra negations. */ |
| michael@0 | 151 | z5 = (tmp10 - tmp12) * ((FAST_FLOAT) 0.382683433); /* c6 */ |
| michael@0 | 152 | z2 = ((FAST_FLOAT) 0.541196100) * tmp10 + z5; /* c2-c6 */ |
| michael@0 | 153 | z4 = ((FAST_FLOAT) 1.306562965) * tmp12 + z5; /* c2+c6 */ |
| michael@0 | 154 | z3 = tmp11 * ((FAST_FLOAT) 0.707106781); /* c4 */ |
| michael@0 | 155 | |
| michael@0 | 156 | z11 = tmp7 + z3; /* phase 5 */ |
| michael@0 | 157 | z13 = tmp7 - z3; |
| michael@0 | 158 | |
| michael@0 | 159 | dataptr[DCTSIZE*5] = z13 + z2; /* phase 6 */ |
| michael@0 | 160 | dataptr[DCTSIZE*3] = z13 - z2; |
| michael@0 | 161 | dataptr[DCTSIZE*1] = z11 + z4; |
| michael@0 | 162 | dataptr[DCTSIZE*7] = z11 - z4; |
| michael@0 | 163 | |
| michael@0 | 164 | dataptr++; /* advance pointer to next column */ |
| michael@0 | 165 | } |
| michael@0 | 166 | } |
| michael@0 | 167 | |
| michael@0 | 168 | #endif /* DCT_FLOAT_SUPPORTED */ |
