4 * Copyright (C) 2007 Marc Hoffman <marc.hoffman@analog.com>
6 * This file is part of FFmpeg.
8 * FFmpeg 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.
13 * FFmpeg 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.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with FFmpeg; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
23 void ff_bfin_fdct (DCTELEM *buf);
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
34 X(m) = sum (x(n) * cos ((2n+1)*m*pi/16))
38 0 --*-------------*0+7---*-----*0+3-------*-*-------------------> 0
40 1 --*-\---------/-*1+6---*-\-/-*1+2-------*-*-------------------> 4
42 2 --*---\-----/---*2+5---*-/-\-*1-2---------------*-*-----------> 2
44 3 --*-----\-/-----*3+4---*-----*0-3---------------*-*-----------> 6
46 4 --*-----/-\-----*3-4------------*-*4+5--*-----*---------------> 1
48 5 --*---/-----\---*2-5---*-*------*=*4-5----\-/------*-*--------> 5
49 / \ X S4,S4 / X S3,-S3
50 6 --*-/---------\-*1-6---*-*------*=*7-6----/-\------*-*--------> 3
52 --*-------------*0-7------------*-*7+6--*-----*---------------> 7
56 Cn = cos(n*pi/8) used throughout the code.
60 R0, R1, R2, R3, R4, R5, R6,R7, P0, P1, P2, P3, P4, P5, A0, A1.
62 I0, I1, I2, I3, B0, B2, B3, M0, M1, L3 registers and LC0.
64 Input - r0 - pointer to start of DCTELEM *block
66 Output - The DCT output coefficients in the DCTELEM *block
69 This code is called from jpeg_encode.
70 R6, R5, R4 if modified should be stored and restored.
73 Performance: (Timer version 0.6.33)
74 Code Size : 240 Bytes.
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
81 -----------------------------------------
82 | Size | Forward DCT | Inverse DCT |
83 -----------------------------------------
84 | 8x8 | 284 Cycles | 311 Cycles |
85 -----------------------------------------
87 Ck = int16(cos(k/16*pi)*32767+.5)/2
93 Sk = int16(sin(k/16*pi)*32767+.5)/2
99 the coefficients are ordered as follows:
106 -----------------------------------------------------------
107 FFMPEG 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:
113 root:/u/ffmpeg/bhead/libavcodec> ./dct-test
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 root:/u/ffmpeg/bhead/libavcodec>
130 #include "config_bfin.h"
132 .section .l1.data.B,"aw",@progbits
135 .short 0x5a82, 0x2d41, 0x187e, 0x3b21, 0x0c7c, 0x3ec5, 0x238e, 0x3537;
137 .section .l1.data.A,"aw",@progbits
143 [--SP] = (R7:4, P5:3); // Push the registers onto the stack.
146 r0 = [P3+dct_coeff@GOT17M4];
148 r0 = [P3+vtmp@GOT17M4];
151 L3 = 16; // L3 is set to 16 to make the coefficient
155 //----------------------------------------------------------------------------
158 * I0, I1, and I2 registers are used to read the input data. I3 register is used
159 * to read the coefficients. P0 and P1 registers are used for writing the output
162 M0 = 12 (X); // All these initializations are used in the
163 M1 = 16 (X); // modification of address offsets.
174 // Prescale the input to get the correct precision.
178 lsetup (.0, .1) LC0 = P3;
180 .0: r1=r0<<3 (v) || r0=[i0++] ;
184 * B0 points to the "in" buffer.
185 * B2 points to "temp" buffer in the first iteration.
188 lsetup (.2, .3) LC0 = P0;
190 I0 = B0; // I0 points to Input Element (0, 0).
191 I1 = B0; // Element 1 and 0 is read in R0.
192 I1 += M0 || R0 = [I0++]; // I1 points to Input Element (0, 6).
193 I2 = I1; // Element 6 is read into R3.H.
194 I2 -= 4 || R3.H = W[I1++]; // I2 points to Input Element (0, 4).
196 I3 = B3; // I3 points to Coefficients.
197 P0 = B2; // P0 points to temporary array Element
199 P1 = B2; // P1 points to temporary array.
200 R7 = [P1++P2] || R2 = [I2++]; // P1 points to temporary array
202 // R7 is a dummy read. X4,X5
204 R3.L = W[I1--]; // X7 is read into R3.L.
205 R1.H = W[I0++]; // X2 is read into R1.H.
209 * X0 = (X0 + X7) / 2.
210 * X1 = (X1 + X6) / 2.
211 * X6 = (X1 - X6) / 2.
212 * X7 = (X0 - X7) / 2.
213 * It reads the data 3 in R1.L.
216 R0 = R0 +|+ R3, R3 = R0 -|- R3 || R1.L = W[I0++] || NOP;
219 * X2 = (X2 + X5) / 2.
220 * X3 = (X3 + X4) / 2.
221 * X4 = (X3 - X4) / 2.
222 * X5 = (X2 - X5) / 2.
223 * R7 = C4 = cos(4*pi/16)
226 R1 = R1 +|+ R2, R2 = R1 -|- R2 (CO) || NOP || R7 = [I3++];
229 * At the end of stage 1 R0 has (1,0), R1 has (2,3), R2 has (4, 5) and
231 * Where the notation (x, y) represents uper/lower half pairs.
240 R0 = R0 +|+ R1, R1 = R0 -|- R1;
242 lsetup (.row0, .row1) LC1 = P2 >> 1; // 1d dct, loops 8x
246 * This is part 2 computation continued.....
247 * A1 = X6 * cos(pi/4)
248 * A0 = X6 * cos(pi/4)
249 * A1 = A1 - X5 * cos(pi/4)
250 * A0 = A0 + X5 * cos(pi/4).
251 * The instruction W[I0] = R3.L is used for packing it to R2.L.
254 A1=R3.H*R7.l, A0=R3.H*R7.l || I1+=M1 || W[I0] = R3.L;
255 R4.H=(A1-=R2.L*R7.l), R4.L=(A0+=R2.L*R7.l) || I2+=M0 || NOP;
257 /* R0 = (X1,X0) R1 = (X2,X3) R4 = (X5, X6). */
260 * A1 = X0 * cos(pi/4)
261 * A0 = X0 * cos(pi/4)
262 * A1 = A1 - X1 * cos(pi/4)
263 * A0 = A0 + X1 * cos(pi/4)
266 A1=R0.L*R7.h, A0=R0.L*R7.h || NOP || R3.H=W[I1++];
267 R5.H=(A1-=R0.H*R7.h),R5.L=(A0+=R0.H*R7.h) || R7=[I3++] || NOP;
270 * A1 = X2 * cos(3pi/8)
271 * A0 = X3 * cos(3pi/8)
272 * A1 = A1 + X3 * cos(pi/8)
273 * A0 = A0 - X2 * cos(pi/8)
275 * R7 = (cos(7pi/8),cos(pi/8))
281 A1=R1.H*R7.L, A0=R1.L*R7.L || W[P0++P3]=R5.L || R2.L=W[I0];
282 R2=R2+|+R4, R4=R2-|-R4 || I0+=4 || R3.L=W[I1--];
283 R6.H=(A1+=R1.L*R7.H),R6.L=(A0 -= R1.H * R7.H) || I0+=4 || R7=[I3++];
285 /* R2 = (X4, X7) R4 = (X5,X6) R5 = (X1, X0) R6 = (X2,X3). */
288 * A1 = X4 * cos(7pi/16)
289 * A0 = X7 * cos(7pi/16)
290 * A1 = A1 + X7 * cos(pi/16)
291 * A0 = A0 - X4 * cos(pi/16)
294 A1=R2.H*R7.L, A0=R2.L*R7.L || W[P0++P3]=R6.H || R0=[I0++];
295 R2.H=(A1+=R2.L*R7.H),R2.L=(A0-=R2.H*R7.H) || W[P0++P3]=R5.H || R7=[I3++];
298 * A1 = X5 * cos(3pi/16)
299 * A0 = X6 * cos(3pi/16)
300 * A1 = A1 + X6 * cos(5pi/16)
301 * A0 = A0 - X5 * cos(5pi/16)
302 * The output values are written.
305 A1=R4.H*R7.H, A0=R4.L*R7.H || W[P0++P2]=R6.L || R1.H=W[I0++];
306 R4.H=(A1+=R4.L*R7.L),R4.L=(A0-=R4.H*R7.L) || W[P0++P4]=R2.L || R1.L=W[I0++];
309 /* Beginning of next stage, **pipelined** + drain and store the
310 rest of the column store. */
312 R0=R0+|+R3,R3=R0-|-R3 || W[P1++P3]=R2.H || R2=[I2++];
313 R1=R1+|+R2,R2=R1-|-R2 (CO) || W[P1++P3]=R4.L || R7=[I3++];
314 .row1: R0=R0+|+R1,R1=R0-|-R1 || W[P1++P5]=R4.H || NOP;
316 // Exchange input with output.
322 (r7:4,p5:3) = [sp++];