4 * Copyright (c) 2001, 2002 Michael Niedermayer <michaelni@gmx.at>
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2 of the License, or (at your option) any later version.
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
20 #include "../dsputil.h"
32 #define C0 23170 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
33 #define C1 22725 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
34 #define C2 21407 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
35 #define C3 19266 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
37 #define C4 16384 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
39 #define C4 16383 //cos(i*M_PI/16)*sqrt(2)*(1<<14) - 0.5
41 #define C5 12873 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
42 #define C6 8867 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
43 #define C7 4520 //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
46 #define COL_SHIFT 20 // 6
48 static const uint64_t __attribute__((aligned(8))) wm1010= 0xFFFF0000FFFF0000ULL;
49 static const uint64_t __attribute__((aligned(8))) d40000= 0x0000000000040000ULL;
51 static const int16_t __attribute__((aligned(8))) coeffs[]= {
52 1<<(ROW_SHIFT-1), 0, 1<<(ROW_SHIFT-1), 0,
53 // 1<<(COL_SHIFT-1), 0, 1<<(COL_SHIFT-1), 0,
54 // 0, 1<<(COL_SHIFT-1-16), 0, 1<<(COL_SHIFT-1-16),
55 1<<(ROW_SHIFT-1), 1, 1<<(ROW_SHIFT-1), 0,
56 // the 1 = ((1<<(COL_SHIFT-1))/C4)<<ROW_SHIFT :)
80 static void unused_var_killer(){
81 int a= wm1010 + d40000;
85 static void inline idctCol (int16_t * col, int16_t *input)
95 int a0, a1, a2, a3, b0, b1, b2, b3;
96 const int C0 = 23170; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
97 const int C1 = 22725; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
98 const int C2 = 21407; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
99 const int C3 = 19266; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
100 const int C4 = 16383; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
101 const int C5 = 12873; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
102 const int C6 = 8867; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
103 const int C7 = 4520; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
105 if( !(col[8*1] | col[8*2] |col[8*3] |col[8*4] |col[8*5] |col[8*6] | col[8*7])) {
106 col[8*0] = col[8*1] = col[8*2] = col[8*3] = col[8*4] =
107 col[8*5] = col[8*6] = col[8*7] = col[8*0]<<3;
111 col[8*0] = input[8*0 + 0];
112 col[8*1] = input[8*2 + 0];
113 col[8*2] = input[8*0 + 1];
114 col[8*3] = input[8*2 + 1];
115 col[8*4] = input[8*4 + 0];
116 col[8*5] = input[8*6 + 0];
117 col[8*6] = input[8*4 + 1];
118 col[8*7] = input[8*6 + 1];
120 a0 = C4*col[8*0] + C2*col[8*2] + C4*col[8*4] + C6*col[8*6] + (1<<(COL_SHIFT-1));
121 a1 = C4*col[8*0] + C6*col[8*2] - C4*col[8*4] - C2*col[8*6] + (1<<(COL_SHIFT-1));
122 a2 = C4*col[8*0] - C6*col[8*2] - C4*col[8*4] + C2*col[8*6] + (1<<(COL_SHIFT-1));
123 a3 = C4*col[8*0] - C2*col[8*2] + C4*col[8*4] - C6*col[8*6] + (1<<(COL_SHIFT-1));
125 b0 = C1*col[8*1] + C3*col[8*3] + C5*col[8*5] + C7*col[8*7];
126 b1 = C3*col[8*1] - C7*col[8*3] - C1*col[8*5] - C5*col[8*7];
127 b2 = C5*col[8*1] - C1*col[8*3] + C7*col[8*5] + C3*col[8*7];
128 b3 = C7*col[8*1] - C5*col[8*3] + C3*col[8*5] - C1*col[8*7];
130 col[8*0] = (a0 + b0) >> COL_SHIFT;
131 col[8*1] = (a1 + b1) >> COL_SHIFT;
132 col[8*2] = (a2 + b2) >> COL_SHIFT;
133 col[8*3] = (a3 + b3) >> COL_SHIFT;
134 col[8*4] = (a3 - b3) >> COL_SHIFT;
135 col[8*5] = (a2 - b2) >> COL_SHIFT;
136 col[8*6] = (a1 - b1) >> COL_SHIFT;
137 col[8*7] = (a0 - b0) >> COL_SHIFT;
140 static void inline idctRow (int16_t * output, int16_t * input)
144 int a0, a1, a2, a3, b0, b1, b2, b3;
145 const int C0 = 23170; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
146 const int C1 = 22725; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
147 const int C2 = 21407; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
148 const int C3 = 19266; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
149 const int C4 = 16383; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
150 const int C5 = 12873; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
151 const int C6 = 8867; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
152 const int C7 = 4520; //cos(i*M_PI/16)*sqrt(2)*(1<<14) + 0.5
163 if( !(row[1] | row[2] |row[3] |row[4] |row[5] |row[6] | row[7]) ) {
164 row[0] = row[1] = row[2] = row[3] = row[4] =
165 row[5] = row[6] = row[7] = row[0]<<3;
177 a0 = C4*row[0] + C2*row[2] + C4*row[4] + C6*row[6] + (1<<(ROW_SHIFT-1));
178 a1 = C4*row[0] + C6*row[2] - C4*row[4] - C2*row[6] + (1<<(ROW_SHIFT-1));
179 a2 = C4*row[0] - C6*row[2] - C4*row[4] + C2*row[6] + (1<<(ROW_SHIFT-1));
180 a3 = C4*row[0] - C2*row[2] + C4*row[4] - C6*row[6] + (1<<(ROW_SHIFT-1));
182 b0 = C1*row[1] + C3*row[3] + C5*row[5] + C7*row[7];
183 b1 = C3*row[1] - C7*row[3] - C1*row[5] - C5*row[7];
184 b2 = C5*row[1] - C1*row[3] + C7*row[5] + C3*row[7];
185 b3 = C7*row[1] - C5*row[3] + C3*row[5] - C1*row[7];
187 row[0] = (a0 + b0) >> ROW_SHIFT;
188 row[1] = (a1 + b1) >> ROW_SHIFT;
189 row[2] = (a2 + b2) >> ROW_SHIFT;
190 row[3] = (a3 + b3) >> ROW_SHIFT;
191 row[4] = (a3 - b3) >> ROW_SHIFT;
192 row[5] = (a2 - b2) >> ROW_SHIFT;
193 row[6] = (a1 - b1) >> ROW_SHIFT;
194 row[7] = (a0 - b0) >> ROW_SHIFT;
207 static inline void idct(int16_t *block)
209 int64_t __attribute__((aligned(8))) align_tmp[16];
210 int16_t * const temp= (int16_t*)align_tmp;
213 #if 0 //Alternative, simpler variant
215 #define ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
216 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
217 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
218 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
219 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
220 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
221 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
222 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
223 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
224 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
225 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
226 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
227 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
228 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
229 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
230 #rounder ", %%mm4 \n\t"\
231 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
232 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
233 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
234 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
235 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
236 #rounder ", %%mm0 \n\t"\
237 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
238 "paddd %%mm0, %%mm0 \n\t" \
239 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
240 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
241 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
242 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
243 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
244 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
245 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
246 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
247 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
248 "psrad $" #shift ", %%mm7 \n\t"\
249 "psrad $" #shift ", %%mm4 \n\t"\
250 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
251 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
252 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
253 "psrad $" #shift ", %%mm1 \n\t"\
254 "psrad $" #shift ", %%mm2 \n\t"\
255 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
256 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
257 "movq %%mm7, " #dst " \n\t"\
258 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
259 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
260 "movq %%mm2, 24+" #dst " \n\t"\
261 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
262 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
263 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
264 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
265 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
266 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
267 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
268 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
269 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
270 "psrad $" #shift ", %%mm2 \n\t"\
271 "psrad $" #shift ", %%mm0 \n\t"\
272 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
273 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
274 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
275 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
276 "psrad $" #shift ", %%mm6 \n\t"\
277 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
278 "movq %%mm2, 8+" #dst " \n\t"\
279 "psrad $" #shift ", %%mm4 \n\t"\
280 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
281 "movq %%mm4, 16+" #dst " \n\t"\
283 #define COL_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
284 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
285 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
286 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
287 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
288 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
289 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
290 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
291 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
292 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
293 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
294 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
295 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
296 #rounder ", %%mm4 \n\t"\
297 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
298 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
299 #rounder ", %%mm0 \n\t"\
300 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
301 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
302 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
303 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
304 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
305 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
306 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
307 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
308 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
309 "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
310 "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
311 "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
312 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
313 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
314 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
315 "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
316 "psrad $" #shift ", %%mm7 \n\t"\
317 "psrad $" #shift ", %%mm4 \n\t"\
318 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
319 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
320 "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
321 "psrad $" #shift ", %%mm0 \n\t"\
322 "psrad $" #shift ", %%mm2 \n\t"\
323 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
324 "movd %%mm7, " #dst " \n\t"\
325 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
326 "movd %%mm0, 16+" #dst " \n\t"\
327 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
328 "movd %%mm2, 96+" #dst " \n\t"\
329 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
330 "movd %%mm4, 112+" #dst " \n\t"\
331 "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
332 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
333 "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
334 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
335 "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
336 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
337 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
338 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
339 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
340 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
341 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
342 "psrad $" #shift ", %%mm2 \n\t"\
343 "psrad $" #shift ", %%mm5 \n\t"\
344 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
345 "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
346 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
347 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
348 "psrad $" #shift ", %%mm6 \n\t"\
349 "psrad $" #shift ", %%mm4 \n\t"\
350 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
351 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
352 "movd %%mm2, 32+" #dst " \n\t"\
353 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
354 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
355 "movd %%mm6, 48+" #dst " \n\t"\
356 "movd %%mm4, 64+" #dst " \n\t"\
357 "movd %%mm5, 80+" #dst " \n\t"\
360 #define DC_COND_ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
361 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
362 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
363 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
364 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
365 "movq "MANGLE(wm1010)", %%mm4 \n\t"\
366 "pand %%mm0, %%mm4 \n\t"\
367 "por %%mm1, %%mm4 \n\t"\
368 "por %%mm2, %%mm4 \n\t"\
369 "por %%mm3, %%mm4 \n\t"\
370 "packssdw %%mm4,%%mm4 \n\t"\
371 "movd %%mm4, %%eax \n\t"\
372 "orl %%eax, %%eax \n\t"\
374 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
375 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
376 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
377 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
378 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
379 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
380 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
381 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
382 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
383 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
384 #rounder ", %%mm4 \n\t"\
385 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
386 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
387 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
388 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
389 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
390 #rounder ", %%mm0 \n\t"\
391 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
392 "paddd %%mm0, %%mm0 \n\t" \
393 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
394 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
395 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
396 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
397 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
398 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
399 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
400 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
401 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
402 "psrad $" #shift ", %%mm7 \n\t"\
403 "psrad $" #shift ", %%mm4 \n\t"\
404 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
405 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
406 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
407 "psrad $" #shift ", %%mm1 \n\t"\
408 "psrad $" #shift ", %%mm2 \n\t"\
409 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
410 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
411 "movq %%mm7, " #dst " \n\t"\
412 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
413 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
414 "movq %%mm2, 24+" #dst " \n\t"\
415 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
416 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
417 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
418 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
419 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
420 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
421 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
422 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
423 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
424 "psrad $" #shift ", %%mm2 \n\t"\
425 "psrad $" #shift ", %%mm0 \n\t"\
426 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
427 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
428 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
429 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
430 "psrad $" #shift ", %%mm6 \n\t"\
431 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
432 "movq %%mm2, 8+" #dst " \n\t"\
433 "psrad $" #shift ", %%mm4 \n\t"\
434 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
435 "movq %%mm4, 16+" #dst " \n\t"\
438 "pslld $16, %%mm0 \n\t"\
439 "#paddd "MANGLE(d40000)", %%mm0 \n\t"\
440 "psrad $13, %%mm0 \n\t"\
441 "packssdw %%mm0, %%mm0 \n\t"\
442 "movq %%mm0, " #dst " \n\t"\
443 "movq %%mm0, 8+" #dst " \n\t"\
444 "movq %%mm0, 16+" #dst " \n\t"\
445 "movq %%mm0, 24+" #dst " \n\t"\
449 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
450 ROW_IDCT( (%0), 8(%0), 16(%0), 24(%0), 0(%1),paddd 8(%2), 11)
451 /*ROW_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1), paddd (%2), 11)
452 ROW_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1), paddd (%2), 11)
453 ROW_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1), paddd (%2), 11)*/
455 DC_COND_ROW_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1),paddd (%2), 11)
456 DC_COND_ROW_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11)
457 DC_COND_ROW_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11)
460 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
461 COL_IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
462 COL_IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
463 COL_IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
464 COL_IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
468 #define DC_COND_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
469 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
470 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
471 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
472 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
473 "movq "MANGLE(wm1010)", %%mm4 \n\t"\
474 "pand %%mm0, %%mm4 \n\t"\
475 "por %%mm1, %%mm4 \n\t"\
476 "por %%mm2, %%mm4 \n\t"\
477 "por %%mm3, %%mm4 \n\t"\
478 "packssdw %%mm4,%%mm4 \n\t"\
479 "movd %%mm4, %%eax \n\t"\
480 "orl %%eax, %%eax \n\t"\
482 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
483 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
484 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
485 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
486 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
487 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
488 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
489 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
490 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
491 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
492 #rounder ", %%mm4 \n\t"\
493 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
494 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
495 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
496 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
497 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
498 #rounder ", %%mm0 \n\t"\
499 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
500 "paddd %%mm0, %%mm0 \n\t" \
501 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
502 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
503 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
504 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
505 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
506 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
507 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
508 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
509 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
510 "psrad $" #shift ", %%mm7 \n\t"\
511 "psrad $" #shift ", %%mm4 \n\t"\
512 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
513 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
514 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
515 "psrad $" #shift ", %%mm1 \n\t"\
516 "psrad $" #shift ", %%mm2 \n\t"\
517 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
518 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
519 "movq %%mm7, " #dst " \n\t"\
520 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
521 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
522 "movq %%mm2, 24+" #dst " \n\t"\
523 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
524 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
525 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
526 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
527 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
528 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
529 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
530 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
531 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
532 "psrad $" #shift ", %%mm2 \n\t"\
533 "psrad $" #shift ", %%mm0 \n\t"\
534 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
535 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
536 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
537 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
538 "psrad $" #shift ", %%mm6 \n\t"\
539 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
540 "movq %%mm2, 8+" #dst " \n\t"\
541 "psrad $" #shift ", %%mm4 \n\t"\
542 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
543 "movq %%mm4, 16+" #dst " \n\t"\
546 "pslld $16, %%mm0 \n\t"\
547 "paddd "MANGLE(d40000)", %%mm0 \n\t"\
548 "psrad $13, %%mm0 \n\t"\
549 "packssdw %%mm0, %%mm0 \n\t"\
550 "movq %%mm0, " #dst " \n\t"\
551 "movq %%mm0, 8+" #dst " \n\t"\
552 "movq %%mm0, 16+" #dst " \n\t"\
553 "movq %%mm0, 24+" #dst " \n\t"\
556 #define Z_COND_IDCT(src0, src4, src1, src5, dst, rounder, shift, bt) \
557 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
558 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
559 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
560 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
561 "movq %%mm0, %%mm4 \n\t"\
562 "por %%mm1, %%mm4 \n\t"\
563 "por %%mm2, %%mm4 \n\t"\
564 "por %%mm3, %%mm4 \n\t"\
565 "packssdw %%mm4,%%mm4 \n\t"\
566 "movd %%mm4, %%eax \n\t"\
567 "orl %%eax, %%eax \n\t"\
569 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
570 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
571 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
572 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
573 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
574 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
575 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
576 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
577 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
578 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
579 #rounder ", %%mm4 \n\t"\
580 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
581 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
582 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
583 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
584 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
585 #rounder ", %%mm0 \n\t"\
586 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
587 "paddd %%mm0, %%mm0 \n\t" \
588 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
589 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
590 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
591 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
592 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
593 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
594 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
595 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
596 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
597 "psrad $" #shift ", %%mm7 \n\t"\
598 "psrad $" #shift ", %%mm4 \n\t"\
599 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
600 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
601 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
602 "psrad $" #shift ", %%mm1 \n\t"\
603 "psrad $" #shift ", %%mm2 \n\t"\
604 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
605 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
606 "movq %%mm7, " #dst " \n\t"\
607 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
608 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
609 "movq %%mm2, 24+" #dst " \n\t"\
610 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
611 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
612 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
613 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
614 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
615 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
616 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
617 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
618 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
619 "psrad $" #shift ", %%mm2 \n\t"\
620 "psrad $" #shift ", %%mm0 \n\t"\
621 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
622 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
623 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
624 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
625 "psrad $" #shift ", %%mm6 \n\t"\
626 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
627 "movq %%mm2, 8+" #dst " \n\t"\
628 "psrad $" #shift ", %%mm4 \n\t"\
629 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
630 "movq %%mm4, 16+" #dst " \n\t"\
632 #define ROW_IDCT(src0, src4, src1, src5, dst, rounder, shift) \
633 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
634 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
635 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
636 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
637 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
638 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
639 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
640 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
641 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
642 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
643 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
644 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
645 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
646 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
647 #rounder ", %%mm4 \n\t"\
648 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
649 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
650 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
651 "movq 56(%2), %%mm5 \n\t" /* C7 C5 C7 C5 */\
652 "pmaddwd %%mm3, %%mm5 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
653 #rounder ", %%mm0 \n\t"\
654 "paddd %%mm0, %%mm1 \n\t" /* A1 a1 */\
655 "paddd %%mm0, %%mm0 \n\t" \
656 "psubd %%mm1, %%mm0 \n\t" /* A2 a2 */\
657 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
658 "paddd %%mm5, %%mm7 \n\t" /* B0 b0 */\
659 "movq 72(%2), %%mm5 \n\t" /* -C5 -C1 -C5 -C1 */\
660 "pmaddwd %%mm3, %%mm5 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
661 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
662 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
663 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
664 "paddd %%mm2, %%mm5 \n\t" /* B1 b1 */\
665 "psrad $" #shift ", %%mm7 \n\t"\
666 "psrad $" #shift ", %%mm4 \n\t"\
667 "movq %%mm1, %%mm2 \n\t" /* A1 a1 */\
668 "paddd %%mm5, %%mm1 \n\t" /* A1+B1 a1+b1 */\
669 "psubd %%mm5, %%mm2 \n\t" /* A1-B1 a1-b1 */\
670 "psrad $" #shift ", %%mm1 \n\t"\
671 "psrad $" #shift ", %%mm2 \n\t"\
672 "packssdw %%mm1, %%mm7 \n\t" /* A1+B1 a1+b1 A0+B0 a0+b0 */\
673 "packssdw %%mm4, %%mm2 \n\t" /* A0-B0 a0-b0 A1-B1 a1-b1 */\
674 "movq %%mm7, " #dst " \n\t"\
675 "movq " #src1 ", %%mm1 \n\t" /* R3 R1 r3 r1 */\
676 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
677 "movq %%mm2, 24+" #dst " \n\t"\
678 "pmaddwd %%mm1, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
679 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
680 "pmaddwd 96(%2), %%mm1 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
681 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
682 "movq %%mm0, %%mm2 \n\t" /* A2 a2 */\
683 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
684 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
685 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
686 "psubd %%mm4, %%mm0 \n\t" /* a2-B2 a2-b2 */\
687 "psrad $" #shift ", %%mm2 \n\t"\
688 "psrad $" #shift ", %%mm0 \n\t"\
689 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
690 "paddd %%mm1, %%mm3 \n\t" /* B3 b3 */\
691 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
692 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
693 "psrad $" #shift ", %%mm6 \n\t"\
694 "packssdw %%mm6, %%mm2 \n\t" /* A3+B3 a3+b3 A2+B2 a2+b2 */\
695 "movq %%mm2, 8+" #dst " \n\t"\
696 "psrad $" #shift ", %%mm4 \n\t"\
697 "packssdw %%mm0, %%mm4 \n\t" /* A2-B2 a2-b2 A3-B3 a3-b3 */\
698 "movq %%mm4, 16+" #dst " \n\t"\
700 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
701 DC_COND_IDCT( 0(%0), 8(%0), 16(%0), 24(%0), 0(%1),paddd 8(%2), 11)
702 Z_COND_IDCT( 32(%0), 40(%0), 48(%0), 56(%0), 32(%1),paddd (%2), 11, 4f)
703 Z_COND_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11, 2f)
704 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 1f)
707 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
708 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
709 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
710 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
711 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
712 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
713 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
714 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
715 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
716 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
717 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
718 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
719 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
720 #rounder ", %%mm4 \n\t"\
721 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
722 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
723 #rounder ", %%mm0 \n\t"\
724 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
725 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
726 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
727 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
728 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
729 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
730 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
731 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
732 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
733 "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
734 "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
735 "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
736 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
737 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
738 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
739 "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
740 "psrad $" #shift ", %%mm7 \n\t"\
741 "psrad $" #shift ", %%mm4 \n\t"\
742 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
743 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
744 "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
745 "psrad $" #shift ", %%mm0 \n\t"\
746 "psrad $" #shift ", %%mm2 \n\t"\
747 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
748 "movd %%mm7, " #dst " \n\t"\
749 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
750 "movd %%mm0, 16+" #dst " \n\t"\
751 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
752 "movd %%mm2, 96+" #dst " \n\t"\
753 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
754 "movd %%mm4, 112+" #dst " \n\t"\
755 "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
756 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
757 "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
758 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
759 "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
760 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
761 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
762 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
763 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
764 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
765 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
766 "psrad $" #shift ", %%mm2 \n\t"\
767 "psrad $" #shift ", %%mm5 \n\t"\
768 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
769 "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
770 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
771 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
772 "psrad $" #shift ", %%mm6 \n\t"\
773 "psrad $" #shift ", %%mm4 \n\t"\
774 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
775 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
776 "movd %%mm2, 32+" #dst " \n\t"\
777 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
778 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
779 "movd %%mm6, 48+" #dst " \n\t"\
780 "movd %%mm4, 64+" #dst " \n\t"\
781 "movd %%mm5, 80+" #dst " \n\t"
784 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
785 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
786 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
787 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
788 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
793 Z_COND_IDCT( 64(%0), 72(%0), 80(%0), 88(%0), 64(%1),paddd (%2), 11, 6f)
794 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 5f)
797 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
798 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
799 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
800 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
801 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
802 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
803 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
804 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
805 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
806 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
807 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
808 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
809 #rounder ", %%mm4 \n\t"\
810 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
811 #rounder ", %%mm0 \n\t"\
812 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
813 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
814 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
815 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
816 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
817 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
818 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
819 "movq 72(%2), %%mm7 \n\t" /* -C5 -C1 -C5 -C1 */\
820 "pmaddwd %%mm3, %%mm7 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
821 "paddd %%mm4, %%mm1 \n\t" /* A0+B0 a0+b0 */\
822 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
823 "psubd %%mm1, %%mm4 \n\t" /* A0-B0 a0-b0 */\
824 "psrad $" #shift ", %%mm1 \n\t"\
825 "psrad $" #shift ", %%mm4 \n\t"\
826 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
827 "paddd %%mm7, %%mm0 \n\t" /* A1+B1 a1+b1 */\
828 "psubd %%mm7, %%mm2 \n\t" /* A1-B1 a1-b1 */\
829 "psrad $" #shift ", %%mm0 \n\t"\
830 "psrad $" #shift ", %%mm2 \n\t"\
831 "packssdw %%mm1, %%mm1 \n\t" /* A0+B0 a0+b0 */\
832 "movd %%mm1, " #dst " \n\t"\
833 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
834 "movd %%mm0, 16+" #dst " \n\t"\
835 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
836 "movd %%mm2, 96+" #dst " \n\t"\
837 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
838 "movd %%mm4, 112+" #dst " \n\t"\
839 "movq 88(%2), %%mm1 \n\t" /* C3 C7 C3 C7 */\
840 "pmaddwd %%mm3, %%mm1 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
841 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
842 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
843 "paddd %%mm1, %%mm2 \n\t" /* A2+B2 a2+b2 */\
844 "psubd %%mm1, %%mm5 \n\t" /* a2-B2 a2-b2 */\
845 "psrad $" #shift ", %%mm2 \n\t"\
846 "psrad $" #shift ", %%mm5 \n\t"\
847 "movq %%mm6, %%mm1 \n\t" /* A3 a3 */\
848 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
849 "psubd %%mm3, %%mm1 \n\t" /* a3-B3 a3-b3 */\
850 "psrad $" #shift ", %%mm6 \n\t"\
851 "psrad $" #shift ", %%mm1 \n\t"\
852 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
853 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
854 "movd %%mm2, 32+" #dst " \n\t"\
855 "packssdw %%mm1, %%mm1 \n\t" /* A3-B3 a3-b3 */\
856 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
857 "movd %%mm6, 48+" #dst " \n\t"\
858 "movd %%mm1, 64+" #dst " \n\t"\
859 "movd %%mm5, 80+" #dst " \n\t"
861 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
862 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
863 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
864 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
865 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
870 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 7f)
873 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
874 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
875 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
876 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
877 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
878 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
879 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
880 #rounder ", %%mm4 \n\t"\
881 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
882 #rounder ", %%mm0 \n\t"\
883 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
884 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
885 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
886 "movq 72(%2), %%mm7 \n\t" /* -C5 -C1 -C5 -C1 */\
887 "pmaddwd %%mm3, %%mm7 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
888 "paddd %%mm4, %%mm1 \n\t" /* A0+B0 a0+b0 */\
889 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
890 "psubd %%mm1, %%mm4 \n\t" /* A0-B0 a0-b0 */\
891 "psrad $" #shift ", %%mm1 \n\t"\
892 "psrad $" #shift ", %%mm4 \n\t"\
893 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
894 "paddd %%mm7, %%mm0 \n\t" /* A1+B1 a1+b1 */\
895 "psubd %%mm7, %%mm2 \n\t" /* A1-B1 a1-b1 */\
896 "psrad $" #shift ", %%mm0 \n\t"\
897 "psrad $" #shift ", %%mm2 \n\t"\
898 "packssdw %%mm1, %%mm1 \n\t" /* A0+B0 a0+b0 */\
899 "movd %%mm1, " #dst " \n\t"\
900 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
901 "movd %%mm0, 16+" #dst " \n\t"\
902 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
903 "movd %%mm2, 96+" #dst " \n\t"\
904 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
905 "movd %%mm4, 112+" #dst " \n\t"\
906 "movq 88(%2), %%mm1 \n\t" /* C3 C7 C3 C7 */\
907 "pmaddwd %%mm3, %%mm1 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
908 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
909 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
910 "paddd %%mm1, %%mm2 \n\t" /* A2+B2 a2+b2 */\
911 "psubd %%mm1, %%mm5 \n\t" /* a2-B2 a2-b2 */\
912 "psrad $" #shift ", %%mm2 \n\t"\
913 "psrad $" #shift ", %%mm5 \n\t"\
914 "movq %%mm6, %%mm1 \n\t" /* A3 a3 */\
915 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
916 "psubd %%mm3, %%mm1 \n\t" /* a3-B3 a3-b3 */\
917 "psrad $" #shift ", %%mm6 \n\t"\
918 "psrad $" #shift ", %%mm1 \n\t"\
919 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
920 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
921 "movd %%mm2, 32+" #dst " \n\t"\
922 "packssdw %%mm1, %%mm1 \n\t" /* A3-B3 a3-b3 */\
923 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
924 "movd %%mm6, 48+" #dst " \n\t"\
925 "movd %%mm1, 64+" #dst " \n\t"\
926 "movd %%mm5, 80+" #dst " \n\t"
929 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
930 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
931 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
932 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
933 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
938 Z_COND_IDCT( 96(%0),104(%0),112(%0),120(%0), 96(%1),paddd (%2), 11, 3f)
941 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
942 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
943 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
944 "movq " #src5 ", %%mm3 \n\t" /* R7 R5 r7 r5 */\
945 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
946 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
947 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
948 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
949 #rounder ", %%mm4 \n\t"\
950 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
951 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
952 #rounder ", %%mm0 \n\t"\
953 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
954 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
955 "movq 56(%2), %%mm1 \n\t" /* C7 C5 C7 C5 */\
956 "pmaddwd %%mm3, %%mm1 \n\t" /* C7R7+C5R5 C7r7+C5r5 */\
957 "pmaddwd 64(%2), %%mm2 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
958 "paddd %%mm1, %%mm7 \n\t" /* B0 b0 */\
959 "movq 72(%2), %%mm1 \n\t" /* -C5 -C1 -C5 -C1 */\
960 "pmaddwd %%mm3, %%mm1 \n\t" /* -C5R7-C1R5 -C5r7-C1r5 */\
961 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
962 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
963 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
964 "paddd %%mm2, %%mm1 \n\t" /* B1 b1 */\
965 "psrad $" #shift ", %%mm7 \n\t"\
966 "psrad $" #shift ", %%mm4 \n\t"\
967 "movq %%mm0, %%mm2 \n\t" /* A1 a1 */\
968 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
969 "psubd %%mm1, %%mm2 \n\t" /* A1-B1 a1-b1 */\
970 "psrad $" #shift ", %%mm0 \n\t"\
971 "psrad $" #shift ", %%mm2 \n\t"\
972 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
973 "movd %%mm7, " #dst " \n\t"\
974 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
975 "movd %%mm0, 16+" #dst " \n\t"\
976 "packssdw %%mm2, %%mm2 \n\t" /* A1-B1 a1-b1 */\
977 "movd %%mm2, 96+" #dst " \n\t"\
978 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
979 "movd %%mm4, 112+" #dst " \n\t"\
980 "movq " #src1 ", %%mm0 \n\t" /* R3 R1 r3 r1 */\
981 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
982 "pmaddwd %%mm0, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
983 "movq 88(%2), %%mm7 \n\t" /* C3 C7 C3 C7 */\
984 "pmaddwd 96(%2), %%mm0 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
985 "pmaddwd %%mm3, %%mm7 \n\t" /* C3R7+C7R5 C3r7+C7r5 */\
986 "movq %%mm5, %%mm2 \n\t" /* A2 a2 */\
987 "pmaddwd 104(%2), %%mm3 \n\t" /* -C1R7+C3R5 -C1r7+C3r5 */\
988 "paddd %%mm7, %%mm4 \n\t" /* B2 b2 */\
989 "paddd %%mm4, %%mm2 \n\t" /* A2+B2 a2+b2 */\
990 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
991 "psrad $" #shift ", %%mm2 \n\t"\
992 "psrad $" #shift ", %%mm5 \n\t"\
993 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
994 "paddd %%mm0, %%mm3 \n\t" /* B3 b3 */\
995 "paddd %%mm3, %%mm6 \n\t" /* A3+B3 a3+b3 */\
996 "psubd %%mm3, %%mm4 \n\t" /* a3-B3 a3-b3 */\
997 "psrad $" #shift ", %%mm6 \n\t"\
998 "psrad $" #shift ", %%mm4 \n\t"\
999 "packssdw %%mm2, %%mm2 \n\t" /* A2+B2 a2+b2 */\
1000 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1001 "movd %%mm2, 32+" #dst " \n\t"\
1002 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
1003 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
1004 "movd %%mm6, 48+" #dst " \n\t"\
1005 "movd %%mm4, 64+" #dst " \n\t"\
1006 "movd %%mm5, 80+" #dst " \n\t"
1008 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
1009 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
1010 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
1011 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
1012 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
1018 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
1019 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
1020 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
1021 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
1022 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1023 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
1024 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1025 #rounder ", %%mm4 \n\t"\
1026 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1027 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
1028 #rounder ", %%mm0 \n\t"\
1029 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
1030 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1031 "movq 64(%2), %%mm3 \n\t"\
1032 "pmaddwd %%mm2, %%mm3 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
1033 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
1034 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
1035 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
1036 "psrad $" #shift ", %%mm7 \n\t"\
1037 "psrad $" #shift ", %%mm4 \n\t"\
1038 "movq %%mm0, %%mm1 \n\t" /* A1 a1 */\
1039 "paddd %%mm3, %%mm0 \n\t" /* A1+B1 a1+b1 */\
1040 "psubd %%mm3, %%mm1 \n\t" /* A1-B1 a1-b1 */\
1041 "psrad $" #shift ", %%mm0 \n\t"\
1042 "psrad $" #shift ", %%mm1 \n\t"\
1043 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
1044 "movd %%mm7, " #dst " \n\t"\
1045 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
1046 "movd %%mm0, 16+" #dst " \n\t"\
1047 "packssdw %%mm1, %%mm1 \n\t" /* A1-B1 a1-b1 */\
1048 "movd %%mm1, 96+" #dst " \n\t"\
1049 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
1050 "movd %%mm4, 112+" #dst " \n\t"\
1051 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
1052 "pmaddwd %%mm2, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
1053 "pmaddwd 96(%2), %%mm2 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
1054 "movq %%mm5, %%mm1 \n\t" /* A2 a2 */\
1055 "paddd %%mm4, %%mm1 \n\t" /* A2+B2 a2+b2 */\
1056 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
1057 "psrad $" #shift ", %%mm1 \n\t"\
1058 "psrad $" #shift ", %%mm5 \n\t"\
1059 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
1060 "paddd %%mm2, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1061 "psubd %%mm2, %%mm4 \n\t" /* a3-B3 a3-b3 */\
1062 "psrad $" #shift ", %%mm6 \n\t"\
1063 "psrad $" #shift ", %%mm4 \n\t"\
1064 "packssdw %%mm1, %%mm1 \n\t" /* A2+B2 a2+b2 */\
1065 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1066 "movd %%mm1, 32+" #dst " \n\t"\
1067 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
1068 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
1069 "movd %%mm6, 48+" #dst " \n\t"\
1070 "movd %%mm4, 64+" #dst " \n\t"\
1071 "movd %%mm5, 80+" #dst " \n\t"
1074 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
1075 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
1076 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
1077 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
1078 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
1084 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
1085 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
1086 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
1087 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
1088 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1089 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
1090 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1091 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
1092 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
1093 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
1094 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
1095 #rounder ", %%mm4 \n\t"\
1096 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1097 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
1098 #rounder ", %%mm0 \n\t"\
1099 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
1100 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1101 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
1102 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
1103 "movq 8+" #src0 ", %%mm2 \n\t" /* R4 R0 r4 r0 */\
1104 "movq 8+" #src4 ", %%mm3 \n\t" /* R6 R2 r6 r2 */\
1105 "movq 16(%2), %%mm1 \n\t" /* C4 C4 C4 C4 */\
1106 "pmaddwd %%mm2, %%mm1 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1107 "movq 24(%2), %%mm7 \n\t" /* -C4 C4 -C4 C4 */\
1108 "pmaddwd %%mm7, %%mm2 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1109 "movq 32(%2), %%mm7 \n\t" /* C6 C2 C6 C2 */\
1110 "pmaddwd %%mm3, %%mm7 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
1111 "pmaddwd 40(%2), %%mm3 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
1112 #rounder ", %%mm1 \n\t"\
1113 "paddd %%mm1, %%mm7 \n\t" /* A0 a0 */\
1114 "paddd %%mm1, %%mm1 \n\t" /* 2C0 2c0 */\
1115 #rounder ", %%mm2 \n\t"\
1116 "psubd %%mm7, %%mm1 \n\t" /* A3 a3 */\
1117 "paddd %%mm2, %%mm3 \n\t" /* A1 a1 */\
1118 "paddd %%mm2, %%mm2 \n\t" /* 2C1 2c1 */\
1119 "psubd %%mm3, %%mm2 \n\t" /* A2 a2 */\
1120 "psrad $" #shift ", %%mm4 \n\t"\
1121 "psrad $" #shift ", %%mm7 \n\t"\
1122 "psrad $" #shift ", %%mm3 \n\t"\
1123 "packssdw %%mm7, %%mm4 \n\t" /* A0 a0 */\
1124 "movq %%mm4, " #dst " \n\t"\
1125 "psrad $" #shift ", %%mm0 \n\t"\
1126 "packssdw %%mm3, %%mm0 \n\t" /* A1 a1 */\
1127 "movq %%mm0, 16+" #dst " \n\t"\
1128 "movq %%mm0, 96+" #dst " \n\t"\
1129 "movq %%mm4, 112+" #dst " \n\t"\
1130 "psrad $" #shift ", %%mm5 \n\t"\
1131 "psrad $" #shift ", %%mm6 \n\t"\
1132 "psrad $" #shift ", %%mm2 \n\t"\
1133 "packssdw %%mm2, %%mm5 \n\t" /* A2-B2 a2-b2 */\
1134 "movq %%mm5, 32+" #dst " \n\t"\
1135 "psrad $" #shift ", %%mm1 \n\t"\
1136 "packssdw %%mm1, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1137 "movq %%mm6, 48+" #dst " \n\t"\
1138 "movq %%mm6, 64+" #dst " \n\t"\
1139 "movq %%mm5, 80+" #dst " \n\t"
1142 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
1143 IDCT( 0(%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
1144 //IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
1145 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
1146 //IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
1153 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
1154 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
1155 "movq " #src4 ", %%mm1 \n\t" /* R6 R2 r6 r2 */\
1156 "movq " #src1 ", %%mm2 \n\t" /* R3 R1 r3 r1 */\
1157 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
1158 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1159 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
1160 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1161 "movq 32(%2), %%mm5 \n\t" /* C6 C2 C6 C2 */\
1162 "pmaddwd %%mm1, %%mm5 \n\t" /* C6R6+C2R2 C6r6+C2r2 */\
1163 "movq 40(%2), %%mm6 \n\t" /* -C2 C6 -C2 C6 */\
1164 "pmaddwd %%mm6, %%mm1 \n\t" /* -C2R6+C6R2 -C2r6+C6r2 */\
1165 #rounder ", %%mm4 \n\t"\
1166 "movq %%mm4, %%mm6 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1167 "movq 48(%2), %%mm7 \n\t" /* C3 C1 C3 C1 */\
1168 #rounder ", %%mm0 \n\t"\
1169 "pmaddwd %%mm2, %%mm7 \n\t" /* C3R3+C1R1 C3r3+C1r1 */\
1170 "paddd %%mm5, %%mm4 \n\t" /* A0 a0 */\
1171 "psubd %%mm5, %%mm6 \n\t" /* A3 a3 */\
1172 "movq %%mm0, %%mm5 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1173 "paddd %%mm1, %%mm0 \n\t" /* A1 a1 */\
1174 "psubd %%mm1, %%mm5 \n\t" /* A2 a2 */\
1175 "movq 64(%2), %%mm1 \n\t"\
1176 "pmaddwd %%mm2, %%mm1 \n\t" /* -C7R3+C3R1 -C7r3+C3r1 */\
1177 "paddd %%mm4, %%mm7 \n\t" /* A0+B0 a0+b0 */\
1178 "paddd %%mm4, %%mm4 \n\t" /* 2A0 2a0 */\
1179 "psubd %%mm7, %%mm4 \n\t" /* A0-B0 a0-b0 */\
1180 "psrad $" #shift ", %%mm7 \n\t"\
1181 "psrad $" #shift ", %%mm4 \n\t"\
1182 "movq %%mm0, %%mm3 \n\t" /* A1 a1 */\
1183 "paddd %%mm1, %%mm0 \n\t" /* A1+B1 a1+b1 */\
1184 "psubd %%mm1, %%mm3 \n\t" /* A1-B1 a1-b1 */\
1185 "psrad $" #shift ", %%mm0 \n\t"\
1186 "psrad $" #shift ", %%mm3 \n\t"\
1187 "packssdw %%mm7, %%mm7 \n\t" /* A0+B0 a0+b0 */\
1188 "movd %%mm7, " #dst " \n\t"\
1189 "packssdw %%mm0, %%mm0 \n\t" /* A1+B1 a1+b1 */\
1190 "movd %%mm0, 16+" #dst " \n\t"\
1191 "packssdw %%mm3, %%mm3 \n\t" /* A1-B1 a1-b1 */\
1192 "movd %%mm3, 96+" #dst " \n\t"\
1193 "packssdw %%mm4, %%mm4 \n\t" /* A0-B0 a0-b0 */\
1194 "movd %%mm4, 112+" #dst " \n\t"\
1195 "movq 80(%2), %%mm4 \n\t" /* -C1 C5 -C1 C5 */\
1196 "pmaddwd %%mm2, %%mm4 \n\t" /* -C1R3+C5R1 -C1r3+C5r1 */\
1197 "pmaddwd 96(%2), %%mm2 \n\t" /* -C5R3+C7R1 -C5r3+C7r1 */\
1198 "movq %%mm5, %%mm3 \n\t" /* A2 a2 */\
1199 "paddd %%mm4, %%mm3 \n\t" /* A2+B2 a2+b2 */\
1200 "psubd %%mm4, %%mm5 \n\t" /* a2-B2 a2-b2 */\
1201 "psrad $" #shift ", %%mm3 \n\t"\
1202 "psrad $" #shift ", %%mm5 \n\t"\
1203 "movq %%mm6, %%mm4 \n\t" /* A3 a3 */\
1204 "paddd %%mm2, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1205 "psubd %%mm2, %%mm4 \n\t" /* a3-B3 a3-b3 */\
1206 "psrad $" #shift ", %%mm6 \n\t"\
1207 "packssdw %%mm3, %%mm3 \n\t" /* A2+B2 a2+b2 */\
1208 "movd %%mm3, 32+" #dst " \n\t"\
1209 "psrad $" #shift ", %%mm4 \n\t"\
1210 "packssdw %%mm6, %%mm6 \n\t" /* A3+B3 a3+b3 */\
1211 "movd %%mm6, 48+" #dst " \n\t"\
1212 "packssdw %%mm4, %%mm4 \n\t" /* A3-B3 a3-b3 */\
1213 "packssdw %%mm5, %%mm5 \n\t" /* A2-B2 a2-b2 */\
1214 "movd %%mm4, 64+" #dst " \n\t"\
1215 "movd %%mm5, 80+" #dst " \n\t"
1218 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
1219 IDCT( (%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
1220 IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
1221 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
1222 IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
1229 #define IDCT(src0, src4, src1, src5, dst, rounder, shift) \
1230 "movq " #src0 ", %%mm0 \n\t" /* R4 R0 r4 r0 */\
1231 "movq 16(%2), %%mm4 \n\t" /* C4 C4 C4 C4 */\
1232 "pmaddwd %%mm0, %%mm4 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1233 "movq 24(%2), %%mm5 \n\t" /* -C4 C4 -C4 C4 */\
1234 "pmaddwd %%mm5, %%mm0 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1235 #rounder ", %%mm4 \n\t"\
1236 #rounder ", %%mm0 \n\t"\
1237 "psrad $" #shift ", %%mm4 \n\t"\
1238 "psrad $" #shift ", %%mm0 \n\t"\
1239 "movq 8+" #src0 ", %%mm2 \n\t" /* R4 R0 r4 r0 */\
1240 "movq 16(%2), %%mm1 \n\t" /* C4 C4 C4 C4 */\
1241 "pmaddwd %%mm2, %%mm1 \n\t" /* C4R4+C4R0 C4r4+C4r0 */\
1242 "movq 24(%2), %%mm7 \n\t" /* -C4 C4 -C4 C4 */\
1243 "pmaddwd %%mm7, %%mm2 \n\t" /* -C4R4+C4R0 -C4r4+C4r0 */\
1244 "movq 32(%2), %%mm7 \n\t" /* C6 C2 C6 C2 */\
1245 #rounder ", %%mm1 \n\t"\
1246 #rounder ", %%mm2 \n\t"\
1247 "psrad $" #shift ", %%mm1 \n\t"\
1248 "packssdw %%mm1, %%mm4 \n\t" /* A0 a0 */\
1249 "movq %%mm4, " #dst " \n\t"\
1250 "psrad $" #shift ", %%mm2 \n\t"\
1251 "packssdw %%mm2, %%mm0 \n\t" /* A1 a1 */\
1252 "movq %%mm0, 16+" #dst " \n\t"\
1253 "movq %%mm0, 96+" #dst " \n\t"\
1254 "movq %%mm4, 112+" #dst " \n\t"\
1255 "movq %%mm0, 32+" #dst " \n\t"\
1256 "movq %%mm4, 48+" #dst " \n\t"\
1257 "movq %%mm4, 64+" #dst " \n\t"\
1258 "movq %%mm0, 80+" #dst " \n\t"
1260 //IDCT( src0, src4, src1, src5, dst, rounder, shift)
1261 IDCT( 0(%1), 64(%1), 32(%1), 96(%1), 0(%0),/nop, 20)
1262 //IDCT( 8(%1), 72(%1), 40(%1), 104(%1), 4(%0),/nop, 20)
1263 IDCT( 16(%1), 80(%1), 48(%1), 112(%1), 8(%0),/nop, 20)
1264 //IDCT( 24(%1), 88(%1), 56(%1), 120(%1), 12(%0),/nop, 20)
1271 00 40 04 44 20 60 24 64
1272 10 30 14 34 50 70 54 74
1273 01 41 03 43 21 61 23 63
1274 11 31 13 33 51 71 53 73
1275 02 42 06 46 22 62 26 66
1276 12 32 16 36 52 72 56 76
1277 05 45 07 47 25 65 27 67
1278 15 35 17 37 55 75 57 77
1281 00 04 10 14 20 24 30 34
1282 40 44 50 54 60 64 70 74
1283 01 03 11 13 21 23 31 33
1284 41 43 51 53 61 63 71 73
1285 02 06 12 16 22 26 32 36
1286 42 46 52 56 62 66 72 76
1287 05 07 15 17 25 27 35 37
1288 45 47 55 57 65 67 75 77
1292 :: "r" (block), "r" (temp), "r" (coeffs)
1297 void ff_simple_idct_mmx(int16_t *block)
1302 //FIXME merge add/put into the idct
1304 void ff_simple_idct_put_mmx(uint8_t *dest, int line_size, DCTELEM *block)
1307 put_pixels_clamped_mmx(block, dest, line_size);
1309 void ff_simple_idct_add_mmx(uint8_t *dest, int line_size, DCTELEM *block)
1312 add_pixels_clamped_mmx(block, dest, line_size);