c923d760e7b0678ca49d984f644f9231f2b44d2b
[libav.git] / libavcodec / bfin / fdct_bfin.S
1 /*
2 * fdct BlackFin
3 *
4 * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
5 *
6 * This file is part of Libav.
7 *
8 * Libav is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
12 *
13 * Libav is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
17 *
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with Libav; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
21 */
22 /*
23 void ff_bfin_fdct (int16_t *buf);
24
25 This implementation works only for 8x8 input. The range of input
26 must be -256 to 255 i.e. 8bit input represented in a 16bit data
27 word. The original data must be sign extended into the 16bit data
28 words.
29
30
31 Chen factorization of
32
33 8
34 X(m) = sum (x(n) * cos ((2n+1)*m*pi/16))
35 n=0
36
37 C4
38 0 --*-------------*0+7---*-----*0+3-------*-*-------------------> 0
39 \ / \ / X S4,S4
40 1 --*-\---------/-*1+6---*-\-/-*1+2-------*-*-------------------> 4
41 \ / \ -C4 C3
42 2 --*---\-----/---*2+5---*-/-\-*1-2---------------*-*-----------> 2
43 \ / / \ X S3,-S3
44 3 --*-----\-/-----*3+4---*-----*0-3---------------*-*-----------> 6
45 / C7 C3
46 4 --*-----/-\-----*3-4------------*-*4+5--*-----*---------------> 1
47 / \ -C4 X \ /S7 C3
48 5 --*---/-----\---*2-5---*-*------*=*4-5----\-/------*-*--------> 5
49 / \ X S4,S4 / X S3,-S3
50 6 --*-/---------\-*1-6---*-*------*=*7-6----/-\------*-*--------> 3
51 / \ C4 X / \-S7 C3
52 --*-------------*0-7------------*-*7+6--*-----*---------------> 7
53 C7
54
55 Notation
56 Cn = cos(n*pi/8) used throughout the code.
57
58
59 Registers used:
60 R0, R1, R2, R3, R4, R5, R6,R7, P0, P1, P2, P3, P4, P5, A0, A1.
61 Other registers used:
62 I0, I1, I2, I3, B0, B2, B3, M0, M1, L3 registers and LC0.
63
64 Input - r0 - pointer to start of int16_t *block
65
66 Output - The DCT output coefficients in the int16_t *block
67
68 Register constraint:
69 This code is called from jpeg_encode.
70 R6, R5, R4 if modified should be stored and restored.
71
72
73 Performance: (Timer version 0.6.33)
74 Code Size : 240 Bytes.
75 Memory Required :
76 Input Matrix : 8 * 8 * 2 Bytes.
77 Coefficients : 16 Bytes
78 Temporary matrix: 8 * 8 * 2 Bytes.
79 Cycle Count :26+{18+8*(14+2S)}*2 where S -> Stalls
80 (7.45 c/pel)
81 -----------------------------------------
82 | Size | Forward DCT | Inverse DCT |
83 -----------------------------------------
84 | 8x8 | 284 Cycles | 311 Cycles |
85 -----------------------------------------
86
87 Ck = int16(cos(k/16*pi)*32767+.5)/2
88 #define C4 23170
89 #define C3 13623
90 #define C6 6270
91 #define C7 3196
92
93 Sk = int16(sin(k/16*pi)*32767+.5)/2
94 #define S4 11585
95 #define S3 9102
96 #define S6 15137
97 #define S7 16069
98
99 the coefficients are ordered as follows:
100 short dct_coef[]
101 C4,S4,
102 C6,S6,
103 C7,S7,
104 S3,C3,
105
106 -----------------------------------------------------------
107 Libav conformance testing results
108 -----------------------------------------------------------
109 dct-test: modified with the following
110 dct_error("BFINfdct", 0, ff_bfin_fdct, fdct, test);
111 produces the following output:
112
113 libavcodec> ./dct-test
114 Libav DCT/IDCT test
115
116 2 -131 -6 -48 -36 33 -83 24
117 34 52 -24 -15 5 92 57 143
118 -67 -43 -1 74 -16 5 -71 32
119 -78 106 92 -34 -38 81 20 -18
120 7 -62 40 2 -15 90 -62 -83
121 -83 1 -104 -13 43 -19 7 11
122 -63 31 12 -29 83 72 21 10
123 -17 -63 -15 73 50 -91 159 -14
124 DCT BFINfdct: err_inf=2 err2=0.16425938 syserr=0.00795000 maxout=2098 blockSumErr=27
125 DCT BFINfdct: 92.1 kdct/s
126 */
127
128 #include "config.h"
129 #include "config_bfin.h"
130
131 #if defined(__FDPIC__) && CONFIG_SRAM
132 .section .l1.data.B,"aw",@progbits
133 #else
134 .data
135 #endif
136 .align 4;
137 dct_coeff:
138 .short 0x5a82, 0x2d41, 0x187e, 0x3b21, 0x0c7c, 0x3ec5, 0x238e, 0x3537;
139
140 #if defined(__FDPIC__) && CONFIG_SRAM
141 .section .l1.data.A,"aw",@progbits
142 #endif
143 .align 4
144 vtmp: .space 128
145
146 .text
147 DEFUN(fdct,mL1,
148 (int16_t *block)):
149 [--SP] = (R7:4, P5:3); // Push the registers onto the stack.
150
151 b0 = r0;
152 RELOC(r0, P3, dct_coeff);
153 b3 = r0;
154 RELOC(r0, P3, vtmp);
155 b2 = r0;
156
157 L3 = 16; // L3 is set to 16 to make the coefficient
158 // array Circular.
159
160
161 //----------------------------------------------------------------------------
162
163 /*
164 * I0, I1, and I2 registers are used to read the input data. I3 register is used
165 * to read the coefficients. P0 and P1 registers are used for writing the output
166 * data.
167 */
168 M0 = 12 (X); // All these initializations are used in the
169 M1 = 16 (X); // modification of address offsets.
170
171 M2 = 128 (X);
172
173 P2 = 16;
174 P3 = 32 (X);
175 P4 = -110 (X);
176 P5 = -62 (X);
177 P0 = 2(X);
178
179
180 // Prescale the input to get the correct precision.
181 i0=b0;
182 i1=b0;
183
184 lsetup (.0, .1) LC0 = P3;
185 r0=[i0++];
186 .0: r1=r0<<3 (v) || r0=[i0++] ;
187 .1: [i1++]=r1;
188
189 /*
190 * B0 points to the "in" buffer.
191 * B2 points to "temp" buffer in the first iteration.
192 */
193
194 lsetup (.2, .3) LC0 = P0;
195 .2:
196 I0 = B0; // I0 points to Input Element (0, 0).
197 I1 = B0; // Element 1 and 0 is read in R0.
198 I1 += M0 || R0 = [I0++]; // I1 points to Input Element (0, 6).
199 I2 = I1; // Element 6 is read into R3.H.
200 I2 -= 4 || R3.H = W[I1++]; // I2 points to Input Element (0, 4).
201
202 I3 = B3; // I3 points to Coefficients.
203 P0 = B2; // P0 points to temporary array Element
204 // (0, 0).
205 P1 = B2; // P1 points to temporary array.
206 R7 = [P1++P2] || R2 = [I2++]; // P1 points to temporary array
207 // Element (1, 0).
208 // R7 is a dummy read. X4,X5
209 // are read into R2.
210 R3.L = W[I1--]; // X7 is read into R3.L.
211 R1.H = W[I0++]; // X2 is read into R1.H.
212
213
214 /*
215 * X0 = (X0 + X7) / 2.
216 * X1 = (X1 + X6) / 2.
217 * X6 = (X1 - X6) / 2.
218 * X7 = (X0 - X7) / 2.
219 * It reads the data 3 in R1.L.
220 */
221
222 R0 = R0 +|+ R3, R3 = R0 -|- R3 || R1.L = W[I0++] || NOP;
223
224 /*
225 * X2 = (X2 + X5) / 2.
226 * X3 = (X3 + X4) / 2.
227 * X4 = (X3 - X4) / 2.
228 * X5 = (X2 - X5) / 2.
229 * R7 = C4 = cos(4*pi/16)
230 */
231
232 R1 = R1 +|+ R2, R2 = R1 -|- R2 (CO) || NOP || R7 = [I3++];
233
234 /*
235 * At the end of stage 1 R0 has (1,0), R1 has (2,3), R2 has (4, 5) and
236 * R3 has (6,7).
237 * Where the notation (x, y) represents uper/lower half pairs.
238 */
239
240 /*
241 * X0 = X0 + X3.
242 * X1 = X1 + X2.
243 * X2 = X1 - X2.
244 * X3 = X0 - X3.
245 */
246 R0 = R0 +|+ R1, R1 = R0 -|- R1;
247
248 lsetup (.row0, .row1) LC1 = P2 >> 1; // 1d dct, loops 8x
249 .row0:
250
251 /*
252 * This is part 2 computation continued.....
253 * A1 = X6 * cos(pi/4)
254 * A0 = X6 * cos(pi/4)
255 * A1 = A1 - X5 * cos(pi/4)
256 * A0 = A0 + X5 * cos(pi/4).
257 * The instruction W[I0] = R3.L is used for packing it to R2.L.
258 */
259
260 A1=R3.H*R7.l, A0=R3.H*R7.l || I1+=M1 || W[I0] = R3.L;
261 R4.H=(A1-=R2.L*R7.l), R4.L=(A0+=R2.L*R7.l) || I2+=M0 || NOP;
262
263 /* R0 = (X1,X0) R1 = (X2,X3) R4 = (X5, X6). */
264
265 /*
266 * A1 = X0 * cos(pi/4)
267 * A0 = X0 * cos(pi/4)
268 * A1 = A1 - X1 * cos(pi/4)
269 * A0 = A0 + X1 * cos(pi/4)
270 * R7 = (C2,C6)
271 */
272 A1=R0.L*R7.h, A0=R0.L*R7.h || NOP || R3.H=W[I1++];
273 R5.H=(A1-=R0.H*R7.h),R5.L=(A0+=R0.H*R7.h) || R7=[I3++] || NOP;
274
275 /*
276 * A1 = X2 * cos(3pi/8)
277 * A0 = X3 * cos(3pi/8)
278 * A1 = A1 + X3 * cos(pi/8)
279 * A0 = A0 - X2 * cos(pi/8)
280 * R3 = cos(pi/4)
281 * R7 = (cos(7pi/8),cos(pi/8))
282 * X4 = X4 + X5.
283 * X5 = X4 - X5.
284 * X6 = X7 - X6.
285 * X7 = X7 + X6.
286 */
287 A1=R1.H*R7.L, A0=R1.L*R7.L || W[P0++P3]=R5.L || R2.L=W[I0];
288 R2=R2+|+R4, R4=R2-|-R4 || I0+=4 || R3.L=W[I1--];
289 R6.H=(A1+=R1.L*R7.H),R6.L=(A0 -= R1.H * R7.H) || I0+=4 || R7=[I3++];
290
291 /* R2 = (X4, X7) R4 = (X5,X6) R5 = (X1, X0) R6 = (X2,X3). */
292
293 /*
294 * A1 = X4 * cos(7pi/16)
295 * A0 = X7 * cos(7pi/16)
296 * A1 = A1 + X7 * cos(pi/16)
297 * A0 = A0 - X4 * cos(pi/16)
298 */
299
300 A1=R2.H*R7.L, A0=R2.L*R7.L || W[P0++P3]=R6.H || R0=[I0++];
301 R2.H=(A1+=R2.L*R7.H),R2.L=(A0-=R2.H*R7.H) || W[P0++P3]=R5.H || R7=[I3++];
302
303 /*
304 * A1 = X5 * cos(3pi/16)
305 * A0 = X6 * cos(3pi/16)
306 * A1 = A1 + X6 * cos(5pi/16)
307 * A0 = A0 - X5 * cos(5pi/16)
308 * The output values are written.
309 */
310
311 A1=R4.H*R7.H, A0=R4.L*R7.H || W[P0++P2]=R6.L || R1.H=W[I0++];
312 R4.H=(A1+=R4.L*R7.L),R4.L=(A0-=R4.H*R7.L) || W[P0++P4]=R2.L || R1.L=W[I0++];
313
314
315 /* Beginning of next stage, **pipelined** + drain and store the
316 rest of the column store. */
317
318 R0=R0+|+R3,R3=R0-|-R3 || W[P1++P3]=R2.H || R2=[I2++];
319 R1=R1+|+R2,R2=R1-|-R2 (CO) || W[P1++P3]=R4.L || R7=[I3++];
320 .row1: R0=R0+|+R1,R1=R0-|-R1 || W[P1++P5]=R4.H || NOP;
321
322 // Exchange input with output.
323 B1 = B0;
324 B0 = B2;
325 .3: B2 = B1;
326
327 L3=0;
328 (r7:4,p5:3) = [sp++];
329 RTS;
330 DEFUN_END(fdct)