1 /*
2 * Floating point AAN DCT
3 * this implementation is based upon the IJG integer AAN DCT (see jfdctfst.c)
4 *
5 * Copyright (c) 2003 Michael Niedermayer <michaelni@gmx.at>
6 * Copyright (c) 2003 Roman Shaposhnik
7 *
8 * Permission to use, copy, modify, and/or distribute this software for any
9 * purpose with or without fee is hereby granted, provided that the above
10 * copyright notice and this permission notice appear in all copies.
11 *
12 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
13 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
14 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
15 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
16 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
17 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
18 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
19 */
20
21 /**
22 * @file
23 * @brief
24 * Floating point AAN DCT
25 * @author Michael Niedermayer <michaelni@gmx.at>
26 */
27
28 #include "faandct.h"
29 #include "libavutil/internal.h"
30 #include "libavutil/libm.h"
31
32 typedef float FLOAT;
33
34 /* numbers generated by arbitrary precision arithmetic followed by truncation
35 to 36 fractional digits (enough for a 128-bit IEEE quad, see /usr/include/math.h
36 for this approach). Unfortunately, long double is not always available correctly,
37 e.g ppc has issues.
38 TODO: add L suffixes when ppc and toolchains sort out their stuff.
39 */
40 #define B0 1.000000000000000000000000000000000000
41 #define B1 0.720959822006947913789091890943021267 // (cos(pi*1/16)sqrt(2))^-1
42 #define B2 0.765366864730179543456919968060797734 // (cos(pi*2/16)sqrt(2))^-1
43 #define B3 0.850430094767256448766702844371412325 // (cos(pi*3/16)sqrt(2))^-1
44 #define B4 1.000000000000000000000000000000000000 // (cos(pi*4/16)sqrt(2))^-1
45 #define B5 1.272758580572833938461007018281767032 // (cos(pi*5/16)sqrt(2))^-1
46 #define B6 1.847759065022573512256366378793576574 // (cos(pi*6/16)sqrt(2))^-1
47 #define B7 3.624509785411551372409941227504289587 // (cos(pi*7/16)sqrt(2))^-1
48
49 #define A1 M_SQRT1_2 // cos(pi*4/16)
50 #define A2 0.54119610014619698435 // cos(pi*6/16)sqrt(2)
51 #define A5 0.38268343236508977170 // cos(pi*6/16)
52 #define A4 1.30656296487637652774 // cos(pi*2/16)sqrt(2)
53
54 static const FLOAT postscale[64]={
55 B0*B0, B0*B1, B0*B2, B0*B3, B0*B4, B0*B5, B0*B6, B0*B7,
56 B1*B0, B1*B1, B1*B2, B1*B3, B1*B4, B1*B5, B1*B6, B1*B7,
57 B2*B0, B2*B1, B2*B2, B2*B3, B2*B4, B2*B5, B2*B6, B2*B7,
58 B3*B0, B3*B1, B3*B2, B3*B3, B3*B4, B3*B5, B3*B6, B3*B7,
59 B4*B0, B4*B1, B4*B2, B4*B3, B4*B4, B4*B5, B4*B6, B4*B7,
60 B5*B0, B5*B1, B5*B2, B5*B3, B5*B4, B5*B5, B5*B6, B5*B7,
61 B6*B0, B6*B1, B6*B2, B6*B3, B6*B4, B6*B5, B6*B6, B6*B7,
62 B7*B0, B7*B1, B7*B2, B7*B3, B7*B4, B7*B5, B7*B6, B7*B7,
63 };
64
row_fdct(FLOAT temp[64],int16_t * data)65 static av_always_inline void row_fdct(FLOAT temp[64], int16_t *data)
66 {
67 FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
68 FLOAT tmp10, tmp11, tmp12, tmp13;
69 FLOAT z2, z4, z11, z13;
70 int i;
71
72 for (i=0; i<8*8; i+=8) {
73 tmp0= data[0 + i] + data[7 + i];
74 tmp7= data[0 + i] - data[7 + i];
75 tmp1= data[1 + i] + data[6 + i];
76 tmp6= data[1 + i] - data[6 + i];
77 tmp2= data[2 + i] + data[5 + i];
78 tmp5= data[2 + i] - data[5 + i];
79 tmp3= data[3 + i] + data[4 + i];
80 tmp4= data[3 + i] - data[4 + i];
81
82 tmp10= tmp0 + tmp3;
83 tmp13= tmp0 - tmp3;
84 tmp11= tmp1 + tmp2;
85 tmp12= tmp1 - tmp2;
86
87 temp[0 + i]= tmp10 + tmp11;
88 temp[4 + i]= tmp10 - tmp11;
89
90 tmp12 += tmp13;
91 tmp12 *= A1;
92 temp[2 + i]= tmp13 + tmp12;
93 temp[6 + i]= tmp13 - tmp12;
94
95 tmp4 += tmp5;
96 tmp5 += tmp6;
97 tmp6 += tmp7;
98
99 z2= tmp4*(A2+A5) - tmp6*A5;
100 z4= tmp6*(A4-A5) + tmp4*A5;
101
102 tmp5*=A1;
103
104 z11= tmp7 + tmp5;
105 z13= tmp7 - tmp5;
106
107 temp[5 + i]= z13 + z2;
108 temp[3 + i]= z13 - z2;
109 temp[1 + i]= z11 + z4;
110 temp[7 + i]= z11 - z4;
111 }
112 }
113
ff_faandct(int16_t * data)114 void ff_faandct(int16_t *data)
115 {
116 FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
117 FLOAT tmp10, tmp11, tmp12, tmp13;
118 FLOAT z2, z4, z11, z13;
119 FLOAT temp[64];
120 int i;
121
122 emms_c();
123
124 row_fdct(temp, data);
125
126 for (i=0; i<8; i++) {
127 tmp0= temp[8*0 + i] + temp[8*7 + i];
128 tmp7= temp[8*0 + i] - temp[8*7 + i];
129 tmp1= temp[8*1 + i] + temp[8*6 + i];
130 tmp6= temp[8*1 + i] - temp[8*6 + i];
131 tmp2= temp[8*2 + i] + temp[8*5 + i];
132 tmp5= temp[8*2 + i] - temp[8*5 + i];
133 tmp3= temp[8*3 + i] + temp[8*4 + i];
134 tmp4= temp[8*3 + i] - temp[8*4 + i];
135
136 tmp10= tmp0 + tmp3;
137 tmp13= tmp0 - tmp3;
138 tmp11= tmp1 + tmp2;
139 tmp12= tmp1 - tmp2;
140
141 data[8*0 + i]= lrintf(postscale[8*0 + i] * (tmp10 + tmp11));
142 data[8*4 + i]= lrintf(postscale[8*4 + i] * (tmp10 - tmp11));
143
144 tmp12 += tmp13;
145 tmp12 *= A1;
146 data[8*2 + i]= lrintf(postscale[8*2 + i] * (tmp13 + tmp12));
147 data[8*6 + i]= lrintf(postscale[8*6 + i] * (tmp13 - tmp12));
148
149 tmp4 += tmp5;
150 tmp5 += tmp6;
151 tmp6 += tmp7;
152
153 z2= tmp4*(A2+A5) - tmp6*A5;
154 z4= tmp6*(A4-A5) + tmp4*A5;
155
156 tmp5*=A1;
157
158 z11= tmp7 + tmp5;
159 z13= tmp7 - tmp5;
160
161 data[8*5 + i]= lrintf(postscale[8*5 + i] * (z13 + z2));
162 data[8*3 + i]= lrintf(postscale[8*3 + i] * (z13 - z2));
163 data[8*1 + i]= lrintf(postscale[8*1 + i] * (z11 + z4));
164 data[8*7 + i]= lrintf(postscale[8*7 + i] * (z11 - z4));
165 }
166 }
167
ff_faandct248(int16_t * data)168 void ff_faandct248(int16_t *data)
169 {
170 FLOAT tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
171 FLOAT tmp10, tmp11, tmp12, tmp13;
172 FLOAT temp[64];
173 int i;
174
175 emms_c();
176
177 row_fdct(temp, data);
178
179 for (i=0; i<8; i++) {
180 tmp0 = temp[8*0 + i] + temp[8*1 + i];
181 tmp1 = temp[8*2 + i] + temp[8*3 + i];
182 tmp2 = temp[8*4 + i] + temp[8*5 + i];
183 tmp3 = temp[8*6 + i] + temp[8*7 + i];
184 tmp4 = temp[8*0 + i] - temp[8*1 + i];
185 tmp5 = temp[8*2 + i] - temp[8*3 + i];
186 tmp6 = temp[8*4 + i] - temp[8*5 + i];
187 tmp7 = temp[8*6 + i] - temp[8*7 + i];
188
189 tmp10 = tmp0 + tmp3;
190 tmp11 = tmp1 + tmp2;
191 tmp12 = tmp1 - tmp2;
192 tmp13 = tmp0 - tmp3;
193
194 data[8*0 + i] = lrintf(postscale[8*0 + i] * (tmp10 + tmp11));
195 data[8*4 + i] = lrintf(postscale[8*4 + i] * (tmp10 - tmp11));
196
197 tmp12 += tmp13;
198 tmp12 *= A1;
199 data[8*2 + i] = lrintf(postscale[8*2 + i] * (tmp13 + tmp12));
200 data[8*6 + i] = lrintf(postscale[8*6 + i] * (tmp13 - tmp12));
201
202 tmp10 = tmp4 + tmp7;
203 tmp11 = tmp5 + tmp6;
204 tmp12 = tmp5 - tmp6;
205 tmp13 = tmp4 - tmp7;
206
207 data[8*1 + i] = lrintf(postscale[8*0 + i] * (tmp10 + tmp11));
208 data[8*5 + i] = lrintf(postscale[8*4 + i] * (tmp10 - tmp11));
209
210 tmp12 += tmp13;
211 tmp12 *= A1;
212 data[8*3 + i] = lrintf(postscale[8*2 + i] * (tmp13 + tmp12));
213 data[8*7 + i] = lrintf(postscale[8*6 + i] * (tmp13 - tmp12));
214 }
215 }
216