Libav
simple_idct.c
Go to the documentation of this file.
1 /*
2  * Simple IDCT MMX
3  *
4  * Copyright (c) 2001, 2002 Michael Niedermayer <michaelni@gmx.at>
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 #include "libavutil/internal.h"
24 #include "libavutil/mem.h"
25 #include "libavutil/x86/asm.h"
26 #include "idctdsp.h"
27 #include "simple_idct.h"
28 
29 #if HAVE_INLINE_ASM
30 
31 /*
32 23170.475006
33 22725.260826
34 21406.727617
35 19265.545870
36 16384.000000
37 12872.826198
38 8866.956905
39 4520.335430
40 */
41 #define C0 23170 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
42 #define C1 22725 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
43 #define C2 21407 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
44 #define C3 19266 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
45 #define C4 16383 //cos(i*M_PI/16)*sqrt(2)*(1<<14) - 0.5
46 #define C5 12873 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
47 #define C6 8867 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
48 #define C7 4520 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
49 
50 #define ROW_SHIFT 11
51 #define COL_SHIFT 20 // 6
52 
53 DECLARE_ASM_CONST(8, uint64_t, wm1010)= 0xFFFF0000FFFF0000ULL;
54 DECLARE_ASM_CONST(8, uint64_t, d40000)= 0x0000000000040000ULL;
55 
56 DECLARE_ALIGNED(8, static const int16_t, coeffs)[]= {
57  1<<(ROW_SHIFT-1), 0, 1<<(ROW_SHIFT-1), 0,
58 // 1<<(COL_SHIFT-1), 0, 1<<(COL_SHIFT-1), 0,
59 // 0, 1<<(COL_SHIFT-1-16), 0, 1<<(COL_SHIFT-1-16),
60  1<<(ROW_SHIFT-1), 1, 1<<(ROW_SHIFT-1), 0,
61  // the 1 = ((1<<(COL_SHIFT-1))/C4)<<ROW_SHIFT :)
62 // 0, 0, 0, 0,
63 // 0, 0, 0, 0,
64 
65  C4, C4, C4, C4,
66  C4, -C4, C4, -C4,
67 
68  C2, C6, C2, C6,
69  C6, -C2, C6, -C2,
70 
71  C1, C3, C1, C3,
72  C5, C7, C5, C7,
73 
74  C3, -C7, C3, -C7,
75 -C1, -C5, -C1, -C5,
76 
77  C5, -C1, C5, -C1,
78  C7, C3, C7, C3,
79 
80  C7, -C5, C7, -C5,
81  C3, -C1, C3, -C1
82 };
83 
84 static inline void idct(int16_t *block)
85 {
86  DECLARE_ALIGNED(8, int64_t, align_tmp)[16];
87  int16_t * const temp= (int16_t*)align_tmp;
88 
89  __asm__ volatile(
90 #if 0 //Alternative, simpler variant
91 
92 #define ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
93  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
94  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
95  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
96  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
97  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
98  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
99  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
100  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
101  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
102  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
103  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
104  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
105  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
106  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
107  #rounder ", %%mm4 \n\t"\
108  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
109  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
110  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
111  "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
112  "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
113  #rounder ", %%mm0 \n\t"\
114  "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
115  "paddd %%mm0, %%mm0 \n\t" \
116  "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
117  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
118  "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
119  "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
120  "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
121  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
122  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
123  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
124  "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
125  "psrad $" #shift ", %%mm7 \n\t"\
126  "psrad $" #shift ", %%mm4 \n\t"\
127  "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
128  "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
129  "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
130  "psrad $" #shift ", %%mm1 \n\t"\
131  "psrad $" #shift ", %%mm2 \n\t"\
132  "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
133  "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
134  "movq %%mm7, " #dst " \n\t"\
135  "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
136  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
137  "movq %%mm2, 24+" #dst " \n\t"\
138  "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
139  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
140  "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
141  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
142  "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
143  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
144  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
145  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
146  "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
147  "psrad $" #shift ", %%mm2 \n\t"\
148  "psrad $" #shift ", %%mm0 \n\t"\
149  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
150  "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
151  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
152  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
153  "psrad $" #shift ", %%mm6 \n\t"\
154  "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
155  "movq %%mm2, 8+" #dst " \n\t"\
156  "psrad $" #shift ", %%mm4 \n\t"\
157  "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
158  "movq %%mm4, 16+" #dst " \n\t"\
159 
160 #define COL_IDCT(src0, src4, src1, src5, dst, shift) \
161  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
162  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
163  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
164  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
165  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
166  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
167  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
168  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
169  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
170  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
171  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
172  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
173  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
174  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
175  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
176  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
177  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
178  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
179  "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
180  "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
181  "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
182  "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
183  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
184  "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
185  "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
186  "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
187  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
188  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
189  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
190  "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
191  "psrad $" #shift ", %%mm7 \n\t"\
192  "psrad $" #shift ", %%mm4 \n\t"\
193  "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
194  "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
195  "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
196  "psrad $" #shift ", %%mm0 \n\t"\
197  "psrad $" #shift ", %%mm2 \n\t"\
198  "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
199  "movd %%mm7, " #dst " \n\t"\
200  "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
201  "movd %%mm0, 16+" #dst " \n\t"\
202  "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
203  "movd %%mm2, 96+" #dst " \n\t"\
204  "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
205  "movd %%mm4, 112+" #dst " \n\t"\
206  "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
207  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
208  "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
209  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
210  "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
211  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
212  "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
213  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
214  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
215  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
216  "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
217  "psrad $" #shift ", %%mm2 \n\t"\
218  "psrad $" #shift ", %%mm5 \n\t"\
219  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
220  "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
221  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
222  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
223  "psrad $" #shift ", %%mm6 \n\t"\
224  "psrad $" #shift ", %%mm4 \n\t"\
225  "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
226  "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
227  "movd %%mm2, 32+" #dst " \n\t"\
228  "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
229  "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
230  "movd %%mm6, 48+" #dst " \n\t"\
231  "movd %%mm4, 64+" #dst " \n\t"\
232  "movd %%mm5, 80+" #dst " \n\t"\
233 
234 
235 #define DC_COND_ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
236  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
237  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
238  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
239  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
240  "movq "MANGLE(wm1010)", %%mm4 \n\t"\
241  "pand %%mm0, %%mm4 \n\t"\
242  "por %%mm1, %%mm4 \n\t"\
243  "por %%mm2, %%mm4 \n\t"\
244  "por %%mm3, %%mm4 \n\t"\
245  "packssdw %%mm4,%%mm4 \n\t"\
246  "movd %%mm4, %%eax \n\t"\
247  "orl %%eax, %%eax \n\t"\
248  "jz 1f \n\t"\
249  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
250  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
251  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
252  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
253  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
254  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
255  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
256  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
257  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
258  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
259  #rounder ", %%mm4 \n\t"\
260  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
261  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
262  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
263  "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
264  "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
265  #rounder ", %%mm0 \n\t"\
266  "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
267  "paddd %%mm0, %%mm0 \n\t" \
268  "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
269  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
270  "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
271  "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
272  "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
273  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
274  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
275  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
276  "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
277  "psrad $" #shift ", %%mm7 \n\t"\
278  "psrad $" #shift ", %%mm4 \n\t"\
279  "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
280  "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
281  "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
282  "psrad $" #shift ", %%mm1 \n\t"\
283  "psrad $" #shift ", %%mm2 \n\t"\
284  "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
285  "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
286  "movq %%mm7, " #dst " \n\t"\
287  "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
288  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
289  "movq %%mm2, 24+" #dst " \n\t"\
290  "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
291  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
292  "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
293  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
294  "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
295  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
296  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
297  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
298  "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
299  "psrad $" #shift ", %%mm2 \n\t"\
300  "psrad $" #shift ", %%mm0 \n\t"\
301  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
302  "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
303  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
304  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
305  "psrad $" #shift ", %%mm6 \n\t"\
306  "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
307  "movq %%mm2, 8+" #dst " \n\t"\
308  "psrad $" #shift ", %%mm4 \n\t"\
309  "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
310  "movq %%mm4, 16+" #dst " \n\t"\
311  "jmp 2f \n\t"\
312  "1: \n\t"\
313  "pslld $16, %%mm0 \n\t"\
314  "#paddd "MANGLE(d40000)", %%mm0 \n\t"\
315  "psrad $13, %%mm0 \n\t"\
316  "packssdw %%mm0, %%mm0 \n\t"\
317  "movq %%mm0, " #dst " \n\t"\
318  "movq %%mm0, 8+" #dst " \n\t"\
319  "movq %%mm0, 16+" #dst " \n\t"\
320  "movq %%mm0, 24+" #dst " \n\t"\
321  "2: \n\t"
322 
323 
324 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
325 ROW_IDCT( (%0), 8(%0), 16(%0), 24(%0), 0(%1),paddd 8(%2), 11)
326 /*ROW_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1), paddd (%2), 11)
327 ROW_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1), paddd (%2), 11)
328 ROW_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1), paddd (%2), 11)*/
329 
330 DC_COND_ROW_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1),paddd (%2), 11)
331 DC_COND_ROW_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11)
332 DC_COND_ROW_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11)
333 
334 
335 //IDCT( src0, src4, src1, src5, dst, shift)
336 COL_IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
337 COL_IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
338 COL_IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
339 COL_IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
340 
341 #else
342 
343 #define DC_COND_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
344  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
345  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
346  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
347  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
348  "movq "MANGLE(wm1010)", %%mm4 \n\t"\
349  "pand %%mm0, %%mm4 \n\t"\
350  "por %%mm1, %%mm4 \n\t"\
351  "por %%mm2, %%mm4 \n\t"\
352  "por %%mm3, %%mm4 \n\t"\
353  "packssdw %%mm4,%%mm4 \n\t"\
354  "movd %%mm4, %%eax \n\t"\
355  "orl %%eax, %%eax \n\t"\
356  "jz 1f \n\t"\
357  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
358  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
359  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
360  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
361  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
362  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
363  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
364  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
365  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
366  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
367  #rounder ", %%mm4 \n\t"\
368  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
369  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
370  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
371  "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
372  "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
373  #rounder ", %%mm0 \n\t"\
374  "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
375  "paddd %%mm0, %%mm0 \n\t" \
376  "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
377  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
378  "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
379  "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
380  "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
381  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
382  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
383  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
384  "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
385  "psrad $" #shift ", %%mm7 \n\t"\
386  "psrad $" #shift ", %%mm4 \n\t"\
387  "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
388  "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
389  "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
390  "psrad $" #shift ", %%mm1 \n\t"\
391  "psrad $" #shift ", %%mm2 \n\t"\
392  "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
393  "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
394  "movq %%mm7, " #dst " \n\t"\
395  "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
396  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
397  "movq %%mm2, 24+" #dst " \n\t"\
398  "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
399  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
400  "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
401  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
402  "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
403  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
404  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
405  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
406  "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
407  "psrad $" #shift ", %%mm2 \n\t"\
408  "psrad $" #shift ", %%mm0 \n\t"\
409  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
410  "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
411  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
412  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
413  "psrad $" #shift ", %%mm6 \n\t"\
414  "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
415  "movq %%mm2, 8+" #dst " \n\t"\
416  "psrad $" #shift ", %%mm4 \n\t"\
417  "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
418  "movq %%mm4, 16+" #dst " \n\t"\
419  "jmp 2f \n\t"\
420  "1: \n\t"\
421  "pslld $16, %%mm0 \n\t"\
422  "paddd "MANGLE(d40000)", %%mm0 \n\t"\
423  "psrad $13, %%mm0 \n\t"\
424  "packssdw %%mm0, %%mm0 \n\t"\
425  "movq %%mm0, " #dst " \n\t"\
426  "movq %%mm0, 8+" #dst " \n\t"\
427  "movq %%mm0, 16+" #dst " \n\t"\
428  "movq %%mm0, 24+" #dst " \n\t"\
429  "2: \n\t"
430 
431 #define Z_COND_IDCT(src0, src4, src1, src5, dst, rounder, shift, bt) \
432  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
433  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
434  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
435  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
436  "movq %%mm0, %%mm4 \n\t"\
437  "por %%mm1, %%mm4 \n\t"\
438  "por %%mm2, %%mm4 \n\t"\
439  "por %%mm3, %%mm4 \n\t"\
440  "packssdw %%mm4,%%mm4 \n\t"\
441  "movd %%mm4, %%eax \n\t"\
442  "orl %%eax, %%eax \n\t"\
443  "jz " #bt " \n\t"\
444  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
445  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
446  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
447  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
448  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
449  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
450  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
451  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
452  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
453  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
454  #rounder ", %%mm4 \n\t"\
455  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
456  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
457  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
458  "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
459  "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
460  #rounder ", %%mm0 \n\t"\
461  "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
462  "paddd %%mm0, %%mm0 \n\t" \
463  "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
464  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
465  "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
466  "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
467  "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
468  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
469  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
470  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
471  "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
472  "psrad $" #shift ", %%mm7 \n\t"\
473  "psrad $" #shift ", %%mm4 \n\t"\
474  "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
475  "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
476  "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
477  "psrad $" #shift ", %%mm1 \n\t"\
478  "psrad $" #shift ", %%mm2 \n\t"\
479  "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
480  "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
481  "movq %%mm7, " #dst " \n\t"\
482  "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
483  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
484  "movq %%mm2, 24+" #dst " \n\t"\
485  "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
486  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
487  "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
488  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
489  "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
490  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
491  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
492  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
493  "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
494  "psrad $" #shift ", %%mm2 \n\t"\
495  "psrad $" #shift ", %%mm0 \n\t"\
496  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
497  "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
498  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
499  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
500  "psrad $" #shift ", %%mm6 \n\t"\
501  "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
502  "movq %%mm2, 8+" #dst " \n\t"\
503  "psrad $" #shift ", %%mm4 \n\t"\
504  "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
505  "movq %%mm4, 16+" #dst " \n\t"\
506 
507 #define ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
508  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
509  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
510  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
511  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
512  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
513  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
514  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
515  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
516  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
517  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
518  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
519  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
520  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
521  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
522  #rounder ", %%mm4 \n\t"\
523  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
524  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
525  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
526  "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
527  "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
528  #rounder ", %%mm0 \n\t"\
529  "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
530  "paddd %%mm0, %%mm0 \n\t" \
531  "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
532  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
533  "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
534  "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
535  "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
536  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
537  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
538  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
539  "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
540  "psrad $" #shift ", %%mm7 \n\t"\
541  "psrad $" #shift ", %%mm4 \n\t"\
542  "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
543  "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
544  "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
545  "psrad $" #shift ", %%mm1 \n\t"\
546  "psrad $" #shift ", %%mm2 \n\t"\
547  "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
548  "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
549  "movq %%mm7, " #dst " \n\t"\
550  "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
551  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
552  "movq %%mm2, 24+" #dst " \n\t"\
553  "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
554  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
555  "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
556  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
557  "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
558  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
559  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
560  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
561  "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
562  "psrad $" #shift ", %%mm2 \n\t"\
563  "psrad $" #shift ", %%mm0 \n\t"\
564  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
565  "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
566  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
567  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
568  "psrad $" #shift ", %%mm6 \n\t"\
569  "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
570  "movq %%mm2, 8+" #dst " \n\t"\
571  "psrad $" #shift ", %%mm4 \n\t"\
572  "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
573  "movq %%mm4, 16+" #dst " \n\t"\
574 
575 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
576 DC_COND_IDCT( 0(%0), 8(%0), 16(%0), 24(%0), 0(%1),paddd 8(%2), 11)
577 Z_COND_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1),paddd (%2), 11, 4f)
578 Z_COND_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11, 2f)
579 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 1f)
580 
581 #undef IDCT
582 #define IDCT(src0, src4, src1, src5, dst, shift) \
583  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
584  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
585  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
586  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
587  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
588  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
589  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
590  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
591  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
592  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
593  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
594  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
595  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
596  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
597  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
598  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
599  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
600  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
601  "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
602  "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
603  "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
604  "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
605  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
606  "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
607  "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
608  "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
609  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
610  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
611  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
612  "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
613  "psrad $" #shift ", %%mm7 \n\t"\
614  "psrad $" #shift ", %%mm4 \n\t"\
615  "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
616  "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
617  "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
618  "psrad $" #shift ", %%mm0 \n\t"\
619  "psrad $" #shift ", %%mm2 \n\t"\
620  "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
621  "movd %%mm7, " #dst " \n\t"\
622  "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
623  "movd %%mm0, 16+" #dst " \n\t"\
624  "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
625  "movd %%mm2, 96+" #dst " \n\t"\
626  "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
627  "movd %%mm4, 112+" #dst " \n\t"\
628  "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
629  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
630  "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
631  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
632  "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
633  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
634  "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
635  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
636  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
637  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
638  "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
639  "psrad $" #shift ", %%mm2 \n\t"\
640  "psrad $" #shift ", %%mm5 \n\t"\
641  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
642  "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
643  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
644  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
645  "psrad $" #shift ", %%mm6 \n\t"\
646  "psrad $" #shift ", %%mm4 \n\t"\
647  "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
648  "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
649  "movd %%mm2, 32+" #dst " \n\t"\
650  "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
651  "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
652  "movd %%mm6, 48+" #dst " \n\t"\
653  "movd %%mm4, 64+" #dst " \n\t"\
654  "movd %%mm5, 80+" #dst " \n\t"
655 
656 
657 //IDCT( src0, src4, src1, src5, dst, shift)
658 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
659 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
660 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
661 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
662  "jmp 9f \n\t"
663 
664  "# .p2align 4 \n\t"\
665  "4: \n\t"
666 Z_COND_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11, 6f)
667 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 5f)
668 
669 #undef IDCT
670 #define IDCT(src0, src4, src1, src5, dst, shift) \
671  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
672  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
673  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
674  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
675  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
676  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
677  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
678  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
679  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
680  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
681  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
682  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
683  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
684  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
685  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
686  "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
687  "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
688  "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
689  "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
690  "movq 72(%2), %%mm7 \n\t" /* -C5 -C1 -C5 -C1 */\
691  "pmaddwd %%mm3, %%mm7 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
692  "paddd %%mm4, %%mm1 \n\t" /* A0+B0 a0+b0 */\
693  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
694  "psubd %%mm1, %%mm4 \n\t" /* A0-B0 a0-b0 */\
695  "psrad $" #shift ", %%mm1 \n\t"\
696  "psrad $" #shift ", %%mm4 \n\t"\
697  "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
698  "paddd %%mm7, %%mm0 \n\t" /* A1+B1 a1+b1 */\
699  "psubd %%mm7, %%mm2 \n\t" /* A1-B1 a1-b1 */\
700  "psrad $" #shift ", %%mm0 \n\t"\
701  "psrad $" #shift ", %%mm2 \n\t"\
702  "packssdw %%mm1, %%mm1 \n\t" /* A0+B0 a0+b0 */\
703  "movd %%mm1, " #dst " \n\t"\
704  "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
705  "movd %%mm0, 16+" #dst " \n\t"\
706  "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
707  "movd %%mm2, 96+" #dst " \n\t"\
708  "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
709  "movd %%mm4, 112+" #dst " \n\t"\
710  "movq 88(%2), %%mm1 \n\t" /* C3 C7 C3 C7 */\
711  "pmaddwd %%mm3, %%mm1 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
712  "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
713  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
714  "paddd %%mm1, %%mm2 \n\t" /* A2+B2 a2+b2 */\
715  "psubd %%mm1, %%mm5 \n\t" /* a2-B2 a2-b2 */\
716  "psrad $" #shift ", %%mm2 \n\t"\
717  "psrad $" #shift ", %%mm5 \n\t"\
718  "movq %%mm6, %%mm1 \n\t" /* A3 a3 */\
719  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
720  "psubd %%mm3, %%mm1 \n\t" /* a3-B3 a3-b3 */\
721  "psrad $" #shift ", %%mm6 \n\t"\
722  "psrad $" #shift ", %%mm1 \n\t"\
723  "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
724  "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
725  "movd %%mm2, 32+" #dst " \n\t"\
726  "packssdw %%mm1, %%mm1 \n\t" /* A3-B3 a3-b3 */\
727  "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
728  "movd %%mm6, 48+" #dst " \n\t"\
729  "movd %%mm1, 64+" #dst " \n\t"\
730  "movd %%mm5, 80+" #dst " \n\t"
731 
732 //IDCT( src0, src4, src1, src5, dst, shift)
733 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
734 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
735 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
736 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
737  "jmp 9f \n\t"
738 
739  "# .p2align 4 \n\t"\
740  "6: \n\t"
741 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 7f)
742 
743 #undef IDCT
744 #define IDCT(src0, src4, src1, src5, dst, shift) \
745  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
746  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
747  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
748  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
749  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
750  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
751  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
752  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
753  "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
754  "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
755  "movq 72(%2), %%mm7 \n\t" /* -C5 -C1 -C5 -C1 */\
756  "pmaddwd %%mm3, %%mm7 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
757  "paddd %%mm4, %%mm1 \n\t" /* A0+B0 a0+b0 */\
758  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
759  "psubd %%mm1, %%mm4 \n\t" /* A0-B0 a0-b0 */\
760  "psrad $" #shift ", %%mm1 \n\t"\
761  "psrad $" #shift ", %%mm4 \n\t"\
762  "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
763  "paddd %%mm7, %%mm0 \n\t" /* A1+B1 a1+b1 */\
764  "psubd %%mm7, %%mm2 \n\t" /* A1-B1 a1-b1 */\
765  "psrad $" #shift ", %%mm0 \n\t"\
766  "psrad $" #shift ", %%mm2 \n\t"\
767  "packssdw %%mm1, %%mm1 \n\t" /* A0+B0 a0+b0 */\
768  "movd %%mm1, " #dst " \n\t"\
769  "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
770  "movd %%mm0, 16+" #dst " \n\t"\
771  "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
772  "movd %%mm2, 96+" #dst " \n\t"\
773  "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
774  "movd %%mm4, 112+" #dst " \n\t"\
775  "movq 88(%2), %%mm1 \n\t" /* C3 C7 C3 C7 */\
776  "pmaddwd %%mm3, %%mm1 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
777  "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
778  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
779  "paddd %%mm1, %%mm2 \n\t" /* A2+B2 a2+b2 */\
780  "psubd %%mm1, %%mm5 \n\t" /* a2-B2 a2-b2 */\
781  "psrad $" #shift ", %%mm2 \n\t"\
782  "psrad $" #shift ", %%mm5 \n\t"\
783  "movq %%mm6, %%mm1 \n\t" /* A3 a3 */\
784  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
785  "psubd %%mm3, %%mm1 \n\t" /* a3-B3 a3-b3 */\
786  "psrad $" #shift ", %%mm6 \n\t"\
787  "psrad $" #shift ", %%mm1 \n\t"\
788  "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
789  "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
790  "movd %%mm2, 32+" #dst " \n\t"\
791  "packssdw %%mm1, %%mm1 \n\t" /* A3-B3 a3-b3 */\
792  "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
793  "movd %%mm6, 48+" #dst " \n\t"\
794  "movd %%mm1, 64+" #dst " \n\t"\
795  "movd %%mm5, 80+" #dst " \n\t"
796 
797 
798 //IDCT( src0, src4, src1, src5, dst, shift)
799 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
800 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
801 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
802 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
803  "jmp 9f \n\t"
804 
805  "# .p2align 4 \n\t"\
806  "2: \n\t"
807 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 3f)
808 
809 #undef IDCT
810 #define IDCT(src0, src4, src1, src5, dst, shift) \
811  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
812  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
813  "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
814  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
815  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
816  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
817  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
818  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
819  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
820  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
821  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
822  "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
823  "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
824  "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
825  "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
826  "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
827  "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
828  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
829  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
830  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
831  "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
832  "psrad $" #shift ", %%mm7 \n\t"\
833  "psrad $" #shift ", %%mm4 \n\t"\
834  "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
835  "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
836  "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
837  "psrad $" #shift ", %%mm0 \n\t"\
838  "psrad $" #shift ", %%mm2 \n\t"\
839  "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
840  "movd %%mm7, " #dst " \n\t"\
841  "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
842  "movd %%mm0, 16+" #dst " \n\t"\
843  "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
844  "movd %%mm2, 96+" #dst " \n\t"\
845  "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
846  "movd %%mm4, 112+" #dst " \n\t"\
847  "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
848  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
849  "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
850  "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
851  "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
852  "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
853  "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
854  "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
855  "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
856  "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
857  "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
858  "psrad $" #shift ", %%mm2 \n\t"\
859  "psrad $" #shift ", %%mm5 \n\t"\
860  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
861  "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
862  "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
863  "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
864  "psrad $" #shift ", %%mm6 \n\t"\
865  "psrad $" #shift ", %%mm4 \n\t"\
866  "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
867  "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
868  "movd %%mm2, 32+" #dst " \n\t"\
869  "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
870  "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
871  "movd %%mm6, 48+" #dst " \n\t"\
872  "movd %%mm4, 64+" #dst " \n\t"\
873  "movd %%mm5, 80+" #dst " \n\t"
874 
875 //IDCT( src0, src4, src1, src5, dst, shift)
876 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
877 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
878 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
879 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
880  "jmp 9f \n\t"
881 
882  "# .p2align 4 \n\t"\
883  "3: \n\t"
884 #undef IDCT
885 #define IDCT(src0, src4, src1, src5, dst, shift) \
886  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
887  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
888  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
889  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
890  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
891  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
892  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
893  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
894  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
895  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
896  "movq 64(%2), %%mm3 \n\t"\
897  "pmaddwd %%mm2, %%mm3 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
898  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
899  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
900  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
901  "psrad $" #shift ", %%mm7 \n\t"\
902  "psrad $" #shift ", %%mm4 \n\t"\
903  "movq %%mm0, %%mm1 \n\t" /* A1 a1 */\
904  "paddd %%mm3, %%mm0 \n\t" /* A1+B1 a1+b1 */\
905  "psubd %%mm3, %%mm1 \n\t" /* A1-B1 a1-b1 */\
906  "psrad $" #shift ", %%mm0 \n\t"\
907  "psrad $" #shift ", %%mm1 \n\t"\
908  "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
909  "movd %%mm7, " #dst " \n\t"\
910  "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
911  "movd %%mm0, 16+" #dst " \n\t"\
912  "packssdw %%mm1, %%mm1 \n\t" /* A1-B1 a1-b1 */\
913  "movd %%mm1, 96+" #dst " \n\t"\
914  "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
915  "movd %%mm4, 112+" #dst " \n\t"\
916  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
917  "pmaddwd %%mm2, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
918  "pmaddwd 96(%2), %%mm2 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
919  "movq %%mm5, %%mm1 \n\t" /* A2 a2 */\
920  "paddd %%mm4, %%mm1 \n\t" /* A2+B2 a2+b2 */\
921  "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
922  "psrad $" #shift ", %%mm1 \n\t"\
923  "psrad $" #shift ", %%mm5 \n\t"\
924  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
925  "paddd %%mm2, %%mm6 \n\t" /* A3+B3 a3+b3 */\
926  "psubd %%mm2, %%mm4 \n\t" /* a3-B3 a3-b3 */\
927  "psrad $" #shift ", %%mm6 \n\t"\
928  "psrad $" #shift ", %%mm4 \n\t"\
929  "packssdw %%mm1, %%mm1 \n\t" /* A2+B2 a2+b2 */\
930  "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
931  "movd %%mm1, 32+" #dst " \n\t"\
932  "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
933  "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
934  "movd %%mm6, 48+" #dst " \n\t"\
935  "movd %%mm4, 64+" #dst " \n\t"\
936  "movd %%mm5, 80+" #dst " \n\t"
937 
938 
939 //IDCT( src0, src4, src1, src5, dst, shift)
940 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
941 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
942 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
943 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
944  "jmp 9f \n\t"
945 
946  "# .p2align 4 \n\t"\
947  "5: \n\t"
948 #undef IDCT
949 #define IDCT(src0, src4, src1, src5, dst, shift) \
950  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
951  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
952  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
953  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
954  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
955  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
956  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
957  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
958  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
959  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
960  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
961  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
962  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
963  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
964  "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
965  "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
966  "movq 8+" #src0 ", %%mm2 \n\t" /* R4 R0 r4 r0 */\
967  "movq 8+" #src4 ", %%mm3 \n\t" /* R6 R2 r6 r2 */\
968  "movq 16(%2), %%mm1 \n\t" /* C4 C4 C4 C4 */\
969  "pmaddwd %%mm2, %%mm1 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
970  "movq 24(%2), %%mm7 \n\t" /* -C4 C4 -C4 C4 */\
971  "pmaddwd %%mm7, %%mm2 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
972  "movq 32(%2), %%mm7 \n\t" /* C6 C2 C6 C2 */\
973  "pmaddwd %%mm3, %%mm7 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
974  "pmaddwd 40(%2), %%mm3 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
975  "paddd %%mm1, %%mm7 \n\t" /* A0 a0 */\
976  "paddd %%mm1, %%mm1 \n\t" /* 2C0 2c0 */\
977  "psubd %%mm7, %%mm1 \n\t" /* A3 a3 */\
978  "paddd %%mm2, %%mm3 \n\t" /* A1 a1 */\
979  "paddd %%mm2, %%mm2 \n\t" /* 2C1 2c1 */\
980  "psubd %%mm3, %%mm2 \n\t" /* A2 a2 */\
981  "psrad $" #shift ", %%mm4 \n\t"\
982  "psrad $" #shift ", %%mm7 \n\t"\
983  "psrad $" #shift ", %%mm3 \n\t"\
984  "packssdw %%mm7, %%mm4 \n\t" /* A0 a0 */\
985  "movq %%mm4, " #dst " \n\t"\
986  "psrad $" #shift ", %%mm0 \n\t"\
987  "packssdw %%mm3, %%mm0 \n\t" /* A1 a1 */\
988  "movq %%mm0, 16+" #dst " \n\t"\
989  "movq %%mm0, 96+" #dst " \n\t"\
990  "movq %%mm4, 112+" #dst " \n\t"\
991  "psrad $" #shift ", %%mm5 \n\t"\
992  "psrad $" #shift ", %%mm6 \n\t"\
993  "psrad $" #shift ", %%mm2 \n\t"\
994  "packssdw %%mm2, %%mm5 \n\t" /* A2-B2 a2-b2 */\
995  "movq %%mm5, 32+" #dst " \n\t"\
996  "psrad $" #shift ", %%mm1 \n\t"\
997  "packssdw %%mm1, %%mm6 \n\t" /* A3+B3 a3+b3 */\
998  "movq %%mm6, 48+" #dst " \n\t"\
999  "movq %%mm6, 64+" #dst " \n\t"\
1000  "movq %%mm5, 80+" #dst " \n\t"
1001 
1002 
1003 //IDCT( src0, src4, src1, src5, dst, shift)
1004 IDCT( 0(%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
1005 //IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
1006 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
1007 //IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
1008  "jmp 9f \n\t"
1009 
1010 
1011  "# .p2align 4 \n\t"\
1012  "1: \n\t"
1013 #undef IDCT
1014 #define IDCT(src0, src4, src1, src5, dst, shift) \
1015  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
1016  "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
1017  "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
1018  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
1019  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1020  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
1021  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1022  "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
1023  "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
1024  "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
1025  "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
1026  "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1027  "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
1028  "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
1029  "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
1030  "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
1031  "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1032  "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
1033  "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
1034  "movq 64(%2), %%mm1 \n\t"\
1035  "pmaddwd %%mm2, %%mm1 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
1036  "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
1037  "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
1038  "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
1039  "psrad $" #shift ", %%mm7 \n\t"\
1040  "psrad $" #shift ", %%mm4 \n\t"\
1041  "movq %%mm0, %%mm3 \n\t" /* A1 a1 */\
1042  "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
1043  "psubd %%mm1, %%mm3 \n\t" /* A1-B1 a1-b1 */\
1044  "psrad $" #shift ", %%mm0 \n\t"\
1045  "psrad $" #shift ", %%mm3 \n\t"\
1046  "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
1047  "movd %%mm7, " #dst " \n\t"\
1048  "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
1049  "movd %%mm0, 16+" #dst " \n\t"\
1050  "packssdw %%mm3, %%mm3 \n\t" /* A1-B1 a1-b1 */\
1051  "movd %%mm3, 96+" #dst " \n\t"\
1052  "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
1053  "movd %%mm4, 112+" #dst " \n\t"\
1054  "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
1055  "pmaddwd %%mm2, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
1056  "pmaddwd 96(%2), %%mm2 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
1057  "movq %%mm5, %%mm3 \n\t" /* A2 a2 */\
1058  "paddd %%mm4, %%mm3 \n\t" /* A2+B2 a2+b2 */\
1059  "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
1060  "psrad $" #shift ", %%mm3 \n\t"\
1061  "psrad $" #shift ", %%mm5 \n\t"\
1062  "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
1063  "paddd %%mm2, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1064  "psubd %%mm2, %%mm4 \n\t" /* a3-B3 a3-b3 */\
1065  "psrad $" #shift ", %%mm6 \n\t"\
1066  "packssdw %%mm3, %%mm3 \n\t" /* A2+B2 a2+b2 */\
1067  "movd %%mm3, 32+" #dst " \n\t"\
1068  "psrad $" #shift ", %%mm4 \n\t"\
1069  "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1070  "movd %%mm6, 48+" #dst " \n\t"\
1071  "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
1072  "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
1073  "movd %%mm4, 64+" #dst " \n\t"\
1074  "movd %%mm5, 80+" #dst " \n\t"
1075 
1076 
1077 //IDCT( src0, src4, src1, src5, dst, shift)
1078 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
1079 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
1080 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
1081 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
1082  "jmp 9f \n\t"
1083 
1084 
1085  "# .p2align 4 \n\t"
1086  "7: \n\t"
1087 #undef IDCT
1088 #define IDCT(src0, src4, src1, src5, dst, shift) \
1089  "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
1090  "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
1091  "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1092  "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
1093  "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1094  "psrad $" #shift ", %%mm4 \n\t"\
1095  "psrad $" #shift ", %%mm0 \n\t"\
1096  "movq 8+" #src0 ", %%mm2 \n\t" /* R4 R0 r4 r0 */\
1097  "movq 16(%2), %%mm1 \n\t" /* C4 C4 C4 C4 */\
1098  "pmaddwd %%mm2, %%mm1 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1099  "movq 24(%2), %%mm7 \n\t" /* -C4 C4 -C4 C4 */\
1100  "pmaddwd %%mm7, %%mm2 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1101  "movq 32(%2), %%mm7 \n\t" /* C6 C2 C6 C2 */\
1102  "psrad $" #shift ", %%mm1 \n\t"\
1103  "packssdw %%mm1, %%mm4 \n\t" /* A0 a0 */\
1104  "movq %%mm4, " #dst " \n\t"\
1105  "psrad $" #shift ", %%mm2 \n\t"\
1106  "packssdw %%mm2, %%mm0 \n\t" /* A1 a1 */\
1107  "movq %%mm0, 16+" #dst " \n\t"\
1108  "movq %%mm0, 96+" #dst " \n\t"\
1109  "movq %%mm4, 112+" #dst " \n\t"\
1110  "movq %%mm0, 32+" #dst " \n\t"\
1111  "movq %%mm4, 48+" #dst " \n\t"\
1112  "movq %%mm4, 64+" #dst " \n\t"\
1113  "movq %%mm0, 80+" #dst " \n\t"
1114 
1115 //IDCT( src0, src4, src1, src5, dst, shift)
1116 IDCT( 0(%1), 64(%1), 32(%1), 96(%1), 0(%0), 20)
1117 //IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0), 20)
1118 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0), 20)
1119 //IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0), 20)
1120 
1121 
1122 #endif
1123 
1124 /*
1125 Input
1126  00 40 04 44 20 60 24 64
1127  10 30 14 34 50 70 54 74
1128  01 41 03 43 21 61 23 63
1129  11 31 13 33 51 71 53 73
1130  02 42 06 46 22 62 26 66
1131  12 32 16 36 52 72 56 76
1132  05 45 07 47 25 65 27 67
1133  15 35 17 37 55 75 57 77
1134 
1135 Temp
1136  00 04 10 14 20 24 30 34
1137  40 44 50 54 60 64 70 74
1138  01 03 11 13 21 23 31 33
1139  41 43 51 53 61 63 71 73
1140  02 06 12 16 22 26 32 36
1141  42 46 52 56 62 66 72 76
1142  05 07 15 17 25 27 35 37
1143  45 47 55 57 65 67 75 77
1144 */
1145 
1146 "9: \n\t"
1147  :: "r" (block), "r" (temp), "r" (coeffs)
1148  : "%eax"
1149  );
1150 }
1151 
1152 void ff_simple_idct_mmx(int16_t *block)
1153 {
1154  idct(block);
1155 }
1156 
1157 //FIXME merge add/put into the idct
1158 
1159 void ff_simple_idct_put_mmx(uint8_t *dest, int line_size, int16_t *block)
1160 {
1161  idct(block);
1162  ff_put_pixels_clamped_mmx(block, dest, line_size);
1163 }
1164 void ff_simple_idct_add_mmx(uint8_t *dest, int line_size, int16_t *block)
1165 {
1166  idct(block);
1167  ff_add_pixels_clamped_mmx(block, dest, line_size);
1168 }
1169 
1170 #endif /* HAVE_INLINE_ASM */
memory handling functions
#define DECLARE_ALIGNED(n, t, v)
Definition: mem.h:58
#define C2
Definition: simple_idct.c:128
#define DECLARE_ASM_CONST(n, t, v)
Definition: mem.h:59
uint8_t
#define C1
Definition: simple_idct.c:127
#define C6
void ff_put_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels, int line_size)
void ff_simple_idct_add_mmx(uint8_t *dest, int line_size, int16_t *block)
common internal API header
#define C5
#define C4
void ff_simple_idct_put_mmx(uint8_t *dest, int line_size, int16_t *block)
void ff_simple_idct_mmx(int16_t *block)
#define C3
Definition: simple_idct.c:129
#define C7
void ff_add_pixels_clamped_mmx(const int16_t *block, uint8_t *pixels, int line_size)
static int idct
Definition: avplay.c:251
static int16_t block[64]
Definition: dct-test.c:88