CoCalc -- mmxfdct.c

GitHub Repository: godotengine/godot
Path: blob/master/thirdparty/libtheora/x86/mmxfdct.c
¹⁰²⁷⁸ views
1
/********************************************************************
2
 *                                                                  *
3
 * THIS FILE IS PART OF THE OggTheora SOFTWARE CODEC SOURCE CODE.   *
4
 * USE, DISTRIBUTION AND REPRODUCTION OF THIS LIBRARY SOURCE IS     *
5
 * GOVERNED BY A BSD-STYLE SOURCE LICENSE INCLUDED WITH THIS SOURCE *
6
 * IN 'COPYING'. PLEASE READ THESE TERMS BEFORE DISTRIBUTING.       *
7
 *                                                                  *
8
 * THE Theora SOURCE CODE IS COPYRIGHT (C) 1999-2006                *
9
 * by the Xiph.Org Foundation https://www.xiph.org/                 *
10
 *                                                                  *
11
 ********************************************************************/
12
/*MMX fDCT implementation for x86_32*/
13
/*$Id: fdct_ses2.c 14579 2008-03-12 06:42:40Z xiphmont $*/
14
#include "x86enc.h"
15
#include "x86zigzag.h"
16

17
#if defined(OC_X86_ASM)
18

19
# define OC_FDCT_STAGE1_8x4 \
20
 "#OC_FDCT_STAGE1_8x4\n\t" \
21
 /*Stage 1:*/ \
22
 /*mm0=t7'=t0-t7*/ \
23
 "psubw %%mm7,%%mm0\n\t" \
24
 "paddw %%mm7,%%mm7\n\t" \
25
 /*mm1=t6'=t1-t6*/ \
26
 "psubw %%mm6,%%mm1\n\t" \
27
 "paddw %%mm6,%%mm6\n\t" \
28
 /*mm2=t5'=t2-t5*/ \
29
 "psubw %%mm5,%%mm2\n\t" \
30
 "paddw %%mm5,%%mm5\n\t" \
31
 /*mm3=t4'=t3-t4*/ \
32
 "psubw %%mm4,%%mm3\n\t" \
33
 "paddw %%mm4,%%mm4\n\t" \
34
 /*mm7=t0'=t0+t7*/ \
35
 "paddw %%mm0,%%mm7\n\t" \
36
 /*mm6=t1'=t1+t6*/ \
37
 "paddw %%mm1,%%mm6\n\t" \
38
 /*mm5=t2'=t2+t5*/ \
39
 "paddw %%mm2,%%mm5\n\t" \
40
 /*mm4=t3'=t3+t4*/ \
41
 "paddw %%mm3,%%mm4\n\t" \
42

43
# define OC_FDCT8x4(_r0,_r1,_r2,_r3,_r4,_r5,_r6,_r7) \
44
 "#OC_FDCT8x4\n\t" \
45
 /*Stage 2:*/ \
46
 /*mm7=t3''=t0'-t3'*/ \
47
 "psubw %%mm4,%%mm7\n\t" \
48
 "paddw %%mm4,%%mm4\n\t" \
49
 /*mm6=t2''=t1'-t2'*/ \
50
 "psubw %%mm5,%%mm6\n\t" \
51
 "movq %%mm7,"_r6"(%[y])\n\t" \
52
 "paddw %%mm5,%%mm5\n\t" \
53
 /*mm1=t5''=t6'-t5'*/ \
54
 "psubw %%mm2,%%mm1\n\t" \
55
 "movq %%mm6,"_r2"(%[y])\n\t" \
56
 /*mm4=t0''=t0'+t3'*/ \
57
 "paddw %%mm7,%%mm4\n\t" \
58
 "paddw %%mm2,%%mm2\n\t" \
59
 /*mm5=t1''=t1'+t2'*/ \
60
 "movq %%mm4,"_r0"(%[y])\n\t" \
61
 "paddw %%mm6,%%mm5\n\t" \
62
 /*mm2=t6''=t6'+t5'*/ \
63
 "paddw %%mm1,%%mm2\n\t" \
64
 "movq %%mm5,"_r4"(%[y])\n\t" \
65
 /*mm0=t7', mm1=t5'', mm2=t6'', mm3=t4'.*/ \
66
 /*mm4, mm5, mm6, mm7 are free.*/ \
67
 /*Stage 3:*/ \
68
 /*mm6={2}x4, mm7={27146,0xB500>>1}x2*/ \
69
 "mov $0x5A806A0A,%[a]\n\t" \
70
 "pcmpeqb %%mm6,%%mm6\n\t" \
71
 "movd %[a],%%mm7\n\t" \
72
 "psrlw $15,%%mm6\n\t" \
73
 "punpckldq %%mm7,%%mm7\n\t" \
74
 "paddw %%mm6,%%mm6\n\t" \
75
 /*mm0=0, m2={-1}x4 \
76
   mm5:mm4=t5''*27146+0xB500*/ \
77
 "movq %%mm1,%%mm4\n\t" \
78
 "movq %%mm1,%%mm5\n\t" \
79
 "punpcklwd %%mm6,%%mm4\n\t" \
80
 "movq %%mm2,"_r3"(%[y])\n\t" \
81
 "pmaddwd %%mm7,%%mm4\n\t" \
82
 "movq %%mm0,"_r7"(%[y])\n\t" \
83
 "punpckhwd %%mm6,%%mm5\n\t" \
84
 "pxor %%mm0,%%mm0\n\t" \
85
 "pmaddwd %%mm7,%%mm5\n\t" \
86
 "pcmpeqb %%mm2,%%mm2\n\t" \
87
 /*mm2=t6'', mm1=t5''+(t5''!=0) \
88
   mm4=(t5''*27146+0xB500>>16)*/ \
89
 "pcmpeqw %%mm1,%%mm0\n\t" \
90
 "psrad $16,%%mm4\n\t" \
91
 "psubw %%mm2,%%mm0\n\t" \
92
 "movq "_r3"(%[y]),%%mm2\n\t" \
93
 "psrad $16,%%mm5\n\t" \
94
 "paddw %%mm0,%%mm1\n\t" \
95
 "packssdw %%mm5,%%mm4\n\t" \
96
 /*mm4=s=(t5''*27146+0xB500>>16)+t5''+(t5''!=0)>>1*/ \
97
 "paddw %%mm1,%%mm4\n\t" \
98
 "movq "_r7"(%[y]),%%mm0\n\t" \
99
 "psraw $1,%%mm4\n\t" \
100
 "movq %%mm3,%%mm1\n\t" \
101
 /*mm3=t4''=t4'+s*/ \
102
 "paddw %%mm4,%%mm3\n\t" \
103
 /*mm1=t5'''=t4'-s*/ \
104
 "psubw %%mm4,%%mm1\n\t" \
105
 /*mm1=0, mm3={-1}x4 \
106
   mm5:mm4=t6''*27146+0xB500*/ \
107
 "movq %%mm2,%%mm4\n\t" \
108
 "movq %%mm2,%%mm5\n\t" \
109
 "punpcklwd %%mm6,%%mm4\n\t" \
110
 "movq %%mm1,"_r5"(%[y])\n\t" \
111
 "pmaddwd %%mm7,%%mm4\n\t" \
112
 "movq %%mm3,"_r1"(%[y])\n\t" \
113
 "punpckhwd %%mm6,%%mm5\n\t" \
114
 "pxor %%mm1,%%mm1\n\t" \
115
 "pmaddwd %%mm7,%%mm5\n\t" \
116
 "pcmpeqb %%mm3,%%mm3\n\t" \
117
 /*mm2=t6''+(t6''!=0), mm4=(t6''*27146+0xB500>>16)*/ \
118
 "psrad $16,%%mm4\n\t" \
119
 "pcmpeqw %%mm2,%%mm1\n\t" \
120
 "psrad $16,%%mm5\n\t" \
121
 "psubw %%mm3,%%mm1\n\t" \
122
 "packssdw %%mm5,%%mm4\n\t" \
123
 "paddw %%mm1,%%mm2\n\t" \
124
 /*mm1=t1'' \
125
   mm4=s=(t6''*27146+0xB500>>16)+t6''+(t6''!=0)>>1*/ \
126
 "paddw %%mm2,%%mm4\n\t" \
127
 "movq "_r4"(%[y]),%%mm1\n\t" \
128
 "psraw $1,%%mm4\n\t" \
129
 "movq %%mm0,%%mm2\n\t" \
130
 /*mm7={54491-0x7FFF,0x7FFF}x2 \
131
   mm0=t7''=t7'+s*/ \
132
 "paddw %%mm4,%%mm0\n\t" \
133
 /*mm2=t6'''=t7'-s*/ \
134
 "psubw %%mm4,%%mm2\n\t" \
135
 /*Stage 4:*/ \
136
 /*mm0=0, mm2=t0'' \
137
   mm5:mm4=t1''*27146+0xB500*/ \
138
 "movq %%mm1,%%mm4\n\t" \
139
 "movq %%mm1,%%mm5\n\t" \
140
 "punpcklwd %%mm6,%%mm4\n\t" \
141
 "movq %%mm2,"_r3"(%[y])\n\t" \
142
 "pmaddwd %%mm7,%%mm4\n\t" \
143
 "movq "_r0"(%[y]),%%mm2\n\t" \
144
 "punpckhwd %%mm6,%%mm5\n\t" \
145
 "movq %%mm0,"_r7"(%[y])\n\t" \
146
 "pmaddwd %%mm7,%%mm5\n\t" \
147
 "pxor %%mm0,%%mm0\n\t" \
148
 /*mm7={27146,0x4000>>1}x2 \
149
   mm0=s=(t1''*27146+0xB500>>16)+t1''+(t1''!=0)*/ \
150
 "psrad $16,%%mm4\n\t" \
151
 "mov $0x20006A0A,%[a]\n\t" \
152
 "pcmpeqw %%mm1,%%mm0\n\t" \
153
 "movd %[a],%%mm7\n\t" \
154
 "psrad $16,%%mm5\n\t" \
155
 "psubw %%mm3,%%mm0\n\t" \
156
 "packssdw %%mm5,%%mm4\n\t" \
157
 "paddw %%mm1,%%mm0\n\t" \
158
 "punpckldq %%mm7,%%mm7\n\t" \
159
 "paddw %%mm4,%%mm0\n\t" \
160
 /*mm6={0x00000E3D}x2 \
161
   mm1=-(t0''==0), mm5:mm4=t0''*27146+0x4000*/ \
162
 "movq %%mm2,%%mm4\n\t" \
163
 "movq %%mm2,%%mm5\n\t" \
164
 "punpcklwd %%mm6,%%mm4\n\t" \
165
 "mov $0x0E3D,%[a]\n\t" \
166
 "pmaddwd %%mm7,%%mm4\n\t" \
167
 "punpckhwd %%mm6,%%mm5\n\t" \
168
 "movd %[a],%%mm6\n\t" \
169
 "pmaddwd %%mm7,%%mm5\n\t" \
170
 "pxor %%mm1,%%mm1\n\t" \
171
 "punpckldq %%mm6,%%mm6\n\t" \
172
 "pcmpeqw %%mm2,%%mm1\n\t" \
173
 /*mm4=r=(t0''*27146+0x4000>>16)+t0''+(t0''!=0)*/ \
174
 "psrad $16,%%mm4\n\t" \
175
 "psubw %%mm3,%%mm1\n\t" \
176
 "psrad $16,%%mm5\n\t" \
177
 "paddw %%mm1,%%mm2\n\t" \
178
 "packssdw %%mm5,%%mm4\n\t" \
179
 "movq "_r5"(%[y]),%%mm1\n\t" \
180
 "paddw %%mm2,%%mm4\n\t" \
181
 /*mm2=t6'', mm0=_y[0]=u=r+s>>1 \
182
   The naive implementation could cause overflow, so we use \
183
    u=(r&s)+((r^s)>>1).*/ \
184
 "movq "_r3"(%[y]),%%mm2\n\t" \
185
 "movq %%mm0,%%mm7\n\t" \
186
 "pxor %%mm4,%%mm0\n\t" \
187
 "pand %%mm4,%%mm7\n\t" \
188
 "psraw $1,%%mm0\n\t" \
189
 "mov $0x7FFF54DC,%[a]\n\t" \
190
 "paddw %%mm7,%%mm0\n\t" \
191
 "movd %[a],%%mm7\n\t" \
192
 /*mm7={54491-0x7FFF,0x7FFF}x2 \
193
   mm4=_y[4]=v=r-u*/ \
194
 "psubw %%mm0,%%mm4\n\t" \
195
 "punpckldq %%mm7,%%mm7\n\t" \
196
 "movq %%mm4,"_r4"(%[y])\n\t" \
197
 /*mm0=0, mm7={36410}x4 \
198
   mm1=(t5'''!=0), mm5:mm4=54491*t5'''+0x0E3D*/ \
199
 "movq %%mm1,%%mm4\n\t" \
200
 "movq %%mm1,%%mm5\n\t" \
201
 "punpcklwd %%mm1,%%mm4\n\t" \
202
 "mov $0x8E3A8E3A,%[a]\n\t" \
203
 "pmaddwd %%mm7,%%mm4\n\t" \
204
 "movq %%mm0,"_r0"(%[y])\n\t" \
205
 "punpckhwd %%mm1,%%mm5\n\t" \
206
 "pxor %%mm0,%%mm0\n\t" \
207
 "pmaddwd %%mm7,%%mm5\n\t" \
208
 "pcmpeqw %%mm0,%%mm1\n\t" \
209
 "movd %[a],%%mm7\n\t" \
210
 "psubw %%mm3,%%mm1\n\t" \
211
 "punpckldq %%mm7,%%mm7\n\t" \
212
 "paddd %%mm6,%%mm4\n\t" \
213
 "paddd %%mm6,%%mm5\n\t" \
214
 /*mm0=0 \
215
   mm3:mm1=36410*t6'''+((t5'''!=0)<<16)*/ \
216
 "movq %%mm2,%%mm6\n\t" \
217
 "movq %%mm2,%%mm3\n\t" \
218
 "pmulhw %%mm7,%%mm6\n\t" \
219
 "paddw %%mm2,%%mm1\n\t" \
220
 "pmullw %%mm7,%%mm3\n\t" \
221
 "pxor %%mm0,%%mm0\n\t" \
222
 "paddw %%mm1,%%mm6\n\t" \
223
 "movq %%mm3,%%mm1\n\t" \
224
 "punpckhwd %%mm6,%%mm3\n\t" \
225
 "punpcklwd %%mm6,%%mm1\n\t" \
226
 /*mm3={-1}x4, mm6={1}x4 \
227
   mm4=_y[5]=u=(54491*t5'''+36410*t6'''+0x0E3D>>16)+(t5'''!=0)*/ \
228
 "paddd %%mm3,%%mm5\n\t" \
229
 "paddd %%mm1,%%mm4\n\t" \
230
 "psrad $16,%%mm5\n\t" \
231
 "pxor %%mm6,%%mm6\n\t" \
232
 "psrad $16,%%mm4\n\t" \
233
 "pcmpeqb %%mm3,%%mm3\n\t" \
234
 "packssdw %%mm5,%%mm4\n\t" \
235
 "psubw %%mm3,%%mm6\n\t" \
236
 /*mm1=t7'', mm7={26568,0x3400}x2 \
237
   mm2=s=t6'''-(36410*u>>16)*/ \
238
 "movq %%mm4,%%mm1\n\t" \
239
 "mov $0x340067C8,%[a]\n\t" \
240
 "pmulhw %%mm7,%%mm4\n\t" \
241
 "movd %[a],%%mm7\n\t" \
242
 "movq %%mm1,"_r5"(%[y])\n\t" \
243
 "punpckldq %%mm7,%%mm7\n\t" \
244
 "paddw %%mm1,%%mm4\n\t" \
245
 "movq "_r7"(%[y]),%%mm1\n\t" \
246
 "psubw %%mm4,%%mm2\n\t" \
247
 /*mm6={0x00007B1B}x2 \
248
   mm0=(s!=0), mm5:mm4=s*26568+0x3400*/ \
249
 "movq %%mm2,%%mm4\n\t" \
250
 "movq %%mm2,%%mm5\n\t" \
251
 "punpcklwd %%mm6,%%mm4\n\t" \
252
 "pcmpeqw %%mm2,%%mm0\n\t" \
253
 "pmaddwd %%mm7,%%mm4\n\t" \
254
 "mov $0x7B1B,%[a]\n\t" \
255
 "punpckhwd %%mm6,%%mm5\n\t" \
256
 "movd %[a],%%mm6\n\t" \
257
 "pmaddwd %%mm7,%%mm5\n\t" \
258
 "psubw %%mm3,%%mm0\n\t" \
259
 "punpckldq %%mm6,%%mm6\n\t" \
260
 /*mm7={64277-0x7FFF,0x7FFF}x2 \
261
   mm2=_y[3]=v=(s*26568+0x3400>>17)+s+(s!=0)*/ \
262
 "psrad $17,%%mm4\n\t" \
263
 "paddw %%mm0,%%mm2\n\t" \
264
 "psrad $17,%%mm5\n\t" \
265
 "mov $0x7FFF7B16,%[a]\n\t" \
266
 "packssdw %%mm5,%%mm4\n\t" \
267
 "movd %[a],%%mm7\n\t" \
268
 "paddw %%mm4,%%mm2\n\t" \
269
 "punpckldq %%mm7,%%mm7\n\t" \
270
 /*mm0=0, mm7={12785}x4 \
271
   mm1=(t7''!=0), mm2=t4'', mm5:mm4=64277*t7''+0x7B1B*/ \
272
 "movq %%mm1,%%mm4\n\t" \
273
 "movq %%mm1,%%mm5\n\t" \
274
 "movq %%mm2,"_r3"(%[y])\n\t" \
275
 "punpcklwd %%mm1,%%mm4\n\t" \
276
 "movq "_r1"(%[y]),%%mm2\n\t" \
277
 "pmaddwd %%mm7,%%mm4\n\t" \
278
 "mov $0x31F131F1,%[a]\n\t" \
279
 "punpckhwd %%mm1,%%mm5\n\t" \
280
 "pxor %%mm0,%%mm0\n\t" \
281
 "pmaddwd %%mm7,%%mm5\n\t" \
282
 "pcmpeqw %%mm0,%%mm1\n\t" \
283
 "movd %[a],%%mm7\n\t" \
284
 "psubw %%mm3,%%mm1\n\t" \
285
 "punpckldq %%mm7,%%mm7\n\t" \
286
 "paddd %%mm6,%%mm4\n\t" \
287
 "paddd %%mm6,%%mm5\n\t" \
288
 /*mm3:mm1=12785*t4'''+((t7''!=0)<<16)*/ \
289
 "movq %%mm2,%%mm6\n\t" \
290
 "movq %%mm2,%%mm3\n\t" \
291
 "pmulhw %%mm7,%%mm6\n\t" \
292
 "pmullw %%mm7,%%mm3\n\t" \
293
 "paddw %%mm1,%%mm6\n\t" \
294
 "movq %%mm3,%%mm1\n\t" \
295
 "punpckhwd %%mm6,%%mm3\n\t" \
296
 "punpcklwd %%mm6,%%mm1\n\t" \
297
 /*mm3={-1}x4, mm6={1}x4 \
298
   mm4=_y[1]=u=(12785*t4'''+64277*t7''+0x7B1B>>16)+(t7''!=0)*/ \
299
 "paddd %%mm3,%%mm5\n\t" \
300
 "paddd %%mm1,%%mm4\n\t" \
301
 "psrad $16,%%mm5\n\t" \
302
 "pxor %%mm6,%%mm6\n\t" \
303
 "psrad $16,%%mm4\n\t" \
304
 "pcmpeqb %%mm3,%%mm3\n\t" \
305
 "packssdw %%mm5,%%mm4\n\t" \
306
 "psubw %%mm3,%%mm6\n\t" \
307
 /*mm1=t3'', mm7={20539,0x3000}x2 \
308
   mm4=s=(12785*u>>16)-t4''*/ \
309
 "movq %%mm4,"_r1"(%[y])\n\t" \
310
 "pmulhw %%mm7,%%mm4\n\t" \
311
 "mov $0x3000503B,%[a]\n\t" \
312
 "movq "_r6"(%[y]),%%mm1\n\t" \
313
 "movd %[a],%%mm7\n\t" \
314
 "psubw %%mm2,%%mm4\n\t" \
315
 "punpckldq %%mm7,%%mm7\n\t" \
316
 /*mm6={0x00006CB7}x2 \
317
   mm0=(s!=0), mm5:mm4=s*20539+0x3000*/ \
318
 "movq %%mm4,%%mm5\n\t" \
319
 "movq %%mm4,%%mm2\n\t" \
320
 "punpcklwd %%mm6,%%mm4\n\t" \
321
 "pcmpeqw %%mm2,%%mm0\n\t" \
322
 "pmaddwd %%mm7,%%mm4\n\t" \
323
 "mov $0x6CB7,%[a]\n\t" \
324
 "punpckhwd %%mm6,%%mm5\n\t" \
325
 "movd %[a],%%mm6\n\t" \
326
 "pmaddwd %%mm7,%%mm5\n\t" \
327
 "psubw %%mm3,%%mm0\n\t" \
328
 "punpckldq %%mm6,%%mm6\n\t" \
329
 /*mm7={60547-0x7FFF,0x7FFF}x2 \
330
   mm2=_y[7]=v=(s*20539+0x3000>>20)+s+(s!=0)*/ \
331
 "psrad $20,%%mm4\n\t" \
332
 "paddw %%mm0,%%mm2\n\t" \
333
 "psrad $20,%%mm5\n\t" \
334
 "mov $0x7FFF6C84,%[a]\n\t" \
335
 "packssdw %%mm5,%%mm4\n\t" \
336
 "movd %[a],%%mm7\n\t" \
337
 "paddw %%mm4,%%mm2\n\t" \
338
 "punpckldq %%mm7,%%mm7\n\t" \
339
 /*mm0=0, mm7={25080}x4 \
340
   mm2=t2'', mm5:mm4=60547*t3''+0x6CB7*/ \
341
 "movq %%mm1,%%mm4\n\t" \
342
 "movq %%mm1,%%mm5\n\t" \
343
 "movq %%mm2,"_r7"(%[y])\n\t" \
344
 "punpcklwd %%mm1,%%mm4\n\t" \
345
 "movq "_r2"(%[y]),%%mm2\n\t" \
346
 "pmaddwd %%mm7,%%mm4\n\t" \
347
 "mov $0x61F861F8,%[a]\n\t" \
348
 "punpckhwd %%mm1,%%mm5\n\t" \
349
 "pxor %%mm0,%%mm0\n\t" \
350
 "pmaddwd %%mm7,%%mm5\n\t" \
351
 "movd %[a],%%mm7\n\t" \
352
 "pcmpeqw %%mm0,%%mm1\n\t" \
353
 "psubw %%mm3,%%mm1\n\t" \
354
 "punpckldq %%mm7,%%mm7\n\t" \
355
 "paddd %%mm6,%%mm4\n\t" \
356
 "paddd %%mm6,%%mm5\n\t" \
357
 /*mm3:mm1=25080*t2''+((t3''!=0)<<16)*/ \
358
 "movq %%mm2,%%mm6\n\t" \
359
 "movq %%mm2,%%mm3\n\t" \
360
 "pmulhw %%mm7,%%mm6\n\t" \
361
 "pmullw %%mm7,%%mm3\n\t" \
362
 "paddw %%mm1,%%mm6\n\t" \
363
 "movq %%mm3,%%mm1\n\t" \
364
 "punpckhwd %%mm6,%%mm3\n\t" \
365
 "punpcklwd %%mm6,%%mm1\n\t" \
366
 /*mm1={-1}x4 \
367
   mm4=u=(25080*t2''+60547*t3''+0x6CB7>>16)+(t3''!=0)*/ \
368
 "paddd %%mm3,%%mm5\n\t" \
369
 "paddd %%mm1,%%mm4\n\t" \
370
 "psrad $16,%%mm5\n\t" \
371
 "mov $0x28005460,%[a]\n\t" \
372
 "psrad $16,%%mm4\n\t" \
373
 "pcmpeqb %%mm1,%%mm1\n\t" \
374
 "packssdw %%mm5,%%mm4\n\t" \
375
 /*mm5={1}x4, mm6=_y[2]=u, mm7={21600,0x2800}x2 \
376
   mm4=s=(25080*u>>16)-t2''*/ \
377
 "movq %%mm4,%%mm6\n\t" \
378
 "pmulhw %%mm7,%%mm4\n\t" \
379
 "pxor %%mm5,%%mm5\n\t" \
380
 "movd %[a],%%mm7\n\t" \
381
 "psubw %%mm1,%%mm5\n\t" \
382
 "punpckldq %%mm7,%%mm7\n\t" \
383
 "psubw %%mm2,%%mm4\n\t" \
384
 /*mm2=s+(s!=0) \
385
   mm4:mm3=s*21600+0x2800*/ \
386
 "movq %%mm4,%%mm3\n\t" \
387
 "movq %%mm4,%%mm2\n\t" \
388
 "punpckhwd %%mm5,%%mm4\n\t" \
389
 "pcmpeqw %%mm2,%%mm0\n\t" \
390
 "pmaddwd %%mm7,%%mm4\n\t" \
391
 "psubw %%mm1,%%mm0\n\t" \
392
 "punpcklwd %%mm5,%%mm3\n\t" \
393
 "paddw %%mm0,%%mm2\n\t" \
394
 "pmaddwd %%mm7,%%mm3\n\t" \
395
 /*mm0=_y[4], mm1=_y[7], mm4=_y[0], mm5=_y[5] \
396
   mm3=_y[6]=v=(s*21600+0x2800>>18)+s+(s!=0)*/ \
397
 "movq "_r4"(%[y]),%%mm0\n\t" \
398
 "psrad $18,%%mm4\n\t" \
399
 "movq "_r5"(%[y]),%%mm5\n\t" \
400
 "psrad $18,%%mm3\n\t" \
401
 "movq "_r7"(%[y]),%%mm1\n\t" \
402
 "packssdw %%mm4,%%mm3\n\t" \
403
 "movq "_r0"(%[y]),%%mm4\n\t" \
404
 "paddw %%mm2,%%mm3\n\t" \
405

406
/*On input, mm4=_y[0], mm6=_y[2], mm0=_y[4], mm5=_y[5], mm3=_y[6], mm1=_y[7].
407
  On output, {_y[4],mm1,mm2,mm3} contains the transpose of _y[4...7] and
408
   {mm4,mm5,mm6,mm7} contains the transpose of _y[0...3].*/
409
# define OC_TRANSPOSE8x4(_r0,_r1,_r2,_r3,_r4,_r5,_r6,_r7) \
410
 "#OC_TRANSPOSE8x4\n\t" \
411
 /*First 4x4 transpose:*/ \
412
 /*mm0 = e3 e2 e1 e0 \
413
   mm5 = f3 f2 f1 f0 \
414
   mm3 = g3 g2 g1 g0 \
415
   mm1 = h3 h2 h1 h0*/ \
416
 "movq %%mm0,%%mm2\n\t" \
417
 "punpcklwd %%mm5,%%mm0\n\t" \
418
 "punpckhwd %%mm5,%%mm2\n\t" \
419
 "movq %%mm3,%%mm5\n\t" \
420
 "punpcklwd %%mm1,%%mm3\n\t" \
421
 "punpckhwd %%mm1,%%mm5\n\t" \
422
 /*mm0 = f1 e1 f0 e0 \
423
   mm2 = f3 e3 f2 e2 \
424
   mm3 = h1 g1 h0 g0 \
425
   mm5 = h3 g3 h2 g2*/ \
426
 "movq %%mm0,%%mm1\n\t" \
427
 "punpckldq %%mm3,%%mm0\n\t" \
428
 "movq %%mm0,"_r4"(%[y])\n\t" \
429
 "punpckhdq %%mm3,%%mm1\n\t" \
430
 "movq "_r1"(%[y]),%%mm0\n\t" \
431
 "movq %%mm2,%%mm3\n\t" \
432
 "punpckldq %%mm5,%%mm2\n\t" \
433
 "punpckhdq %%mm5,%%mm3\n\t" \
434
 "movq "_r3"(%[y]),%%mm5\n\t" \
435
 /*_y[4] = h0 g0 f0 e0 \
436
    mm1  = h1 g1 f1 e1 \
437
    mm2  = h2 g2 f2 e2 \
438
    mm3  = h3 g3 f3 e3*/ \
439
 /*Second 4x4 transpose:*/ \
440
 /*mm4 = a3 a2 a1 a0 \
441
   mm0 = b3 b2 b1 b0 \
442
   mm6 = c3 c2 c1 c0 \
443
   mm5 = d3 d2 d1 d0*/ \
444
 "movq %%mm4,%%mm7\n\t" \
445
 "punpcklwd %%mm0,%%mm4\n\t" \
446
 "punpckhwd %%mm0,%%mm7\n\t" \
447
 "movq %%mm6,%%mm0\n\t" \
448
 "punpcklwd %%mm5,%%mm6\n\t" \
449
 "punpckhwd %%mm5,%%mm0\n\t" \
450
 /*mm4 = b1 a1 b0 a0 \
451
   mm7 = b3 a3 b2 a2 \
452
   mm6 = d1 c1 d0 c0 \
453
   mm0 = d3 c3 d2 c2*/ \
454
 "movq %%mm4,%%mm5\n\t" \
455
 "punpckldq %%mm6,%%mm4\n\t" \
456
 "punpckhdq %%mm6,%%mm5\n\t" \
457
 "movq %%mm7,%%mm6\n\t" \
458
 "punpckhdq %%mm0,%%mm7\n\t" \
459
 "punpckldq %%mm0,%%mm6\n\t" \
460
 /*mm4 = d0 c0 b0 a0 \
461
   mm5 = d1 c1 b1 a1 \
462
   mm6 = d2 c2 b2 a2 \
463
   mm7 = d3 c3 b3 a3*/ \
464

465
/*MMX implementation of the fDCT.*/
466
void oc_enc_fdct8x8_mmxext(ogg_int16_t _y[64],const ogg_int16_t _x[64]){
467
  OC_ALIGN8(ogg_int16_t buf[64]);
468
  ptrdiff_t   a;
469
  __asm__ __volatile__(
470
    /*Add two extra bits of working precision to improve accuracy; any more and
471
       we could overflow.*/
472
    /*We also add biases to correct for some systematic error that remains in
473
       the full fDCT->iDCT round trip.*/
474
    "movq 0x00(%[x]),%%mm0\n\t"
475
    "movq 0x10(%[x]),%%mm1\n\t"
476
    "movq 0x20(%[x]),%%mm2\n\t"
477
    "movq 0x30(%[x]),%%mm3\n\t"
478
    "pcmpeqb %%mm4,%%mm4\n\t"
479
    "pxor %%mm7,%%mm7\n\t"
480
    "movq %%mm0,%%mm5\n\t"
481
    "psllw $2,%%mm0\n\t"
482
    "pcmpeqw %%mm7,%%mm5\n\t"
483
    "movq 0x70(%[x]),%%mm7\n\t"
484
    "psllw $2,%%mm1\n\t"
485
    "psubw %%mm4,%%mm5\n\t"
486
    "psllw $2,%%mm2\n\t"
487
    "mov $1,%[a]\n\t"
488
    "pslld $16,%%mm5\n\t"
489
    "movd %[a],%%mm6\n\t"
490
    "psllq $16,%%mm5\n\t"
491
    "mov $0x10001,%[a]\n\t"
492
    "psllw $2,%%mm3\n\t"
493
    "movd %[a],%%mm4\n\t"
494
    "punpckhwd %%mm6,%%mm5\n\t"
495
    "psubw %%mm6,%%mm1\n\t"
496
    "movq 0x60(%[x]),%%mm6\n\t"
497
    "paddw %%mm5,%%mm0\n\t"
498
    "movq 0x50(%[x]),%%mm5\n\t"
499
    "paddw %%mm4,%%mm0\n\t"
500
    "movq 0x40(%[x]),%%mm4\n\t"
501
    /*We inline stage1 of the transform here so we can get better instruction
502
       scheduling with the shifts.*/
503
    /*mm0=t7'=t0-t7*/
504
    "psllw $2,%%mm7\n\t"
505
    "psubw %%mm7,%%mm0\n\t"
506
    "psllw $2,%%mm6\n\t"
507
    "paddw %%mm7,%%mm7\n\t"
508
    /*mm1=t6'=t1-t6*/
509
    "psllw $2,%%mm5\n\t"
510
    "psubw %%mm6,%%mm1\n\t"
511
    "psllw $2,%%mm4\n\t"
512
    "paddw %%mm6,%%mm6\n\t"
513
    /*mm2=t5'=t2-t5*/
514
    "psubw %%mm5,%%mm2\n\t"
515
    "paddw %%mm5,%%mm5\n\t"
516
    /*mm3=t4'=t3-t4*/
517
    "psubw %%mm4,%%mm3\n\t"
518
    "paddw %%mm4,%%mm4\n\t"
519
    /*mm7=t0'=t0+t7*/
520
    "paddw %%mm0,%%mm7\n\t"
521
    /*mm6=t1'=t1+t6*/
522
    "paddw %%mm1,%%mm6\n\t"
523
    /*mm5=t2'=t2+t5*/
524
    "paddw %%mm2,%%mm5\n\t"
525
    /*mm4=t3'=t3+t4*/
526
    "paddw %%mm3,%%mm4\n\t"
527
    OC_FDCT8x4("0x00","0x10","0x20","0x30","0x40","0x50","0x60","0x70")
528
    OC_TRANSPOSE8x4("0x00","0x10","0x20","0x30","0x40","0x50","0x60","0x70")
529
    /*Swap out this 8x4 block for the next one.*/
530
    "movq 0x08(%[x]),%%mm0\n\t"
531
    "movq %%mm7,0x30(%[y])\n\t"
532
    "movq 0x78(%[x]),%%mm7\n\t"
533
    "movq %%mm1,0x50(%[y])\n\t"
534
    "movq 0x18(%[x]),%%mm1\n\t"
535
    "movq %%mm6,0x20(%[y])\n\t"
536
    "movq 0x68(%[x]),%%mm6\n\t"
537
    "movq %%mm2,0x60(%[y])\n\t"
538
    "movq 0x28(%[x]),%%mm2\n\t"
539
    "movq %%mm5,0x10(%[y])\n\t"
540
    "movq 0x58(%[x]),%%mm5\n\t"
541
    "movq %%mm3,0x70(%[y])\n\t"
542
    "movq 0x38(%[x]),%%mm3\n\t"
543
    /*And increase its working precision, too.*/
544
    "psllw $2,%%mm0\n\t"
545
    "movq %%mm4,0x00(%[y])\n\t"
546
    "psllw $2,%%mm7\n\t"
547
    "movq 0x48(%[x]),%%mm4\n\t"
548
    /*We inline stage1 of the transform here so we can get better instruction
549
       scheduling with the shifts.*/
550
    /*mm0=t7'=t0-t7*/
551
    "psubw %%mm7,%%mm0\n\t"
552
    "psllw $2,%%mm1\n\t"
553
    "paddw %%mm7,%%mm7\n\t"
554
    "psllw $2,%%mm6\n\t"
555
    /*mm1=t6'=t1-t6*/
556
    "psubw %%mm6,%%mm1\n\t"
557
    "psllw $2,%%mm2\n\t"
558
    "paddw %%mm6,%%mm6\n\t"
559
    "psllw $2,%%mm5\n\t"
560
    /*mm2=t5'=t2-t5*/
561
    "psubw %%mm5,%%mm2\n\t"
562
    "psllw $2,%%mm3\n\t"
563
    "paddw %%mm5,%%mm5\n\t"
564
    "psllw $2,%%mm4\n\t"
565
    /*mm3=t4'=t3-t4*/
566
    "psubw %%mm4,%%mm3\n\t"
567
    "paddw %%mm4,%%mm4\n\t"
568
    /*mm7=t0'=t0+t7*/
569
    "paddw %%mm0,%%mm7\n\t"
570
    /*mm6=t1'=t1+t6*/
571
    "paddw %%mm1,%%mm6\n\t"
572
    /*mm5=t2'=t2+t5*/
573
    "paddw %%mm2,%%mm5\n\t"
574
    /*mm4=t3'=t3+t4*/
575
    "paddw %%mm3,%%mm4\n\t"
576
    OC_FDCT8x4("0x08","0x18","0x28","0x38","0x48","0x58","0x68","0x78")
577
    OC_TRANSPOSE8x4("0x08","0x18","0x28","0x38","0x48","0x58","0x68","0x78")
578
    /*Here the first 4x4 block of output from the last transpose is the second
579
       4x4 block of input for the next transform.
580
      We have cleverly arranged that it already be in the appropriate place,
581
       so we only have to do half the stores and loads.*/
582
    "movq 0x00(%[y]),%%mm0\n\t"
583
    "movq %%mm1,0x58(%[y])\n\t"
584
    "movq 0x10(%[y]),%%mm1\n\t"
585
    "movq %%mm2,0x68(%[y])\n\t"
586
    "movq 0x20(%[y]),%%mm2\n\t"
587
    "movq %%mm3,0x78(%[y])\n\t"
588
    "movq 0x30(%[y]),%%mm3\n\t"
589
    OC_FDCT_STAGE1_8x4
590
    OC_FDCT8x4("0x00","0x10","0x20","0x30","0x08","0x18","0x28","0x38")
591
    /*mm2={-2}x4*/
592
    "pcmpeqw %%mm2,%%mm2\n\t"
593
    "paddw %%mm2,%%mm2\n\t"
594
    /*Round and store the results (no transpose).*/
595
    "movq 0x10(%[y]),%%mm7\n\t"
596
    "psubw %%mm2,%%mm4\n\t"
597
    "psubw %%mm2,%%mm6\n\t"
598
    "psraw $2,%%mm4\n\t"
599
    "psubw %%mm2,%%mm0\n\t"
600
    "movq %%mm4,"OC_MEM_OFFS(0x00,buf)"\n\t"
601
    "movq 0x30(%[y]),%%mm4\n\t"
602
    "psraw $2,%%mm6\n\t"
603
    "psubw %%mm2,%%mm5\n\t"
604
    "movq %%mm6,"OC_MEM_OFFS(0x20,buf)"\n\t"
605
    "psraw $2,%%mm0\n\t"
606
    "psubw %%mm2,%%mm3\n\t"
607
    "movq %%mm0,"OC_MEM_OFFS(0x40,buf)"\n\t"
608
    "psraw $2,%%mm5\n\t"
609
    "psubw %%mm2,%%mm1\n\t"
610
    "movq %%mm5,"OC_MEM_OFFS(0x50,buf)"\n\t"
611
    "psraw $2,%%mm3\n\t"
612
    "psubw %%mm2,%%mm7\n\t"
613
    "movq %%mm3,"OC_MEM_OFFS(0x60,buf)"\n\t"
614
    "psraw $2,%%mm1\n\t"
615
    "psubw %%mm2,%%mm4\n\t"
616
    "movq %%mm1,"OC_MEM_OFFS(0x70,buf)"\n\t"
617
    "psraw $2,%%mm7\n\t"
618
    "movq %%mm7,"OC_MEM_OFFS(0x10,buf)"\n\t"
619
    "psraw $2,%%mm4\n\t"
620
    "movq %%mm4,"OC_MEM_OFFS(0x30,buf)"\n\t"
621
    /*Load the next block.*/
622
    "movq 0x40(%[y]),%%mm0\n\t"
623
    "movq 0x78(%[y]),%%mm7\n\t"
624
    "movq 0x50(%[y]),%%mm1\n\t"
625
    "movq 0x68(%[y]),%%mm6\n\t"
626
    "movq 0x60(%[y]),%%mm2\n\t"
627
    "movq 0x58(%[y]),%%mm5\n\t"
628
    "movq 0x70(%[y]),%%mm3\n\t"
629
    "movq 0x48(%[y]),%%mm4\n\t"
630
    OC_FDCT_STAGE1_8x4
631
    OC_FDCT8x4("0x40","0x50","0x60","0x70","0x48","0x58","0x68","0x78")
632
    /*mm2={-2}x4*/
633
    "pcmpeqw %%mm2,%%mm2\n\t"
634
    "paddw %%mm2,%%mm2\n\t"
635
    /*Round and store the results (no transpose).*/
636
    "movq 0x50(%[y]),%%mm7\n\t"
637
    "psubw %%mm2,%%mm4\n\t"
638
    "psubw %%mm2,%%mm6\n\t"
639
    "psraw $2,%%mm4\n\t"
640
    "psubw %%mm2,%%mm0\n\t"
641
    "movq %%mm4,"OC_MEM_OFFS(0x08,buf)"\n\t"
642
    "movq 0x70(%[y]),%%mm4\n\t"
643
    "psraw $2,%%mm6\n\t"
644
    "psubw %%mm2,%%mm5\n\t"
645
    "movq %%mm6,"OC_MEM_OFFS(0x28,buf)"\n\t"
646
    "psraw $2,%%mm0\n\t"
647
    "psubw %%mm2,%%mm3\n\t"
648
    "movq %%mm0,"OC_MEM_OFFS(0x48,buf)"\n\t"
649
    "psraw $2,%%mm5\n\t"
650
    "psubw %%mm2,%%mm1\n\t"
651
    "movq %%mm5,"OC_MEM_OFFS(0x58,buf)"\n\t"
652
    "psraw $2,%%mm3\n\t"
653
    "psubw %%mm2,%%mm7\n\t"
654
    "movq %%mm3,"OC_MEM_OFFS(0x68,buf)"\n\t"
655
    "psraw $2,%%mm1\n\t"
656
    "psubw %%mm2,%%mm4\n\t"
657
    "movq %%mm1,"OC_MEM_OFFS(0x78,buf)"\n\t"
658
    "psraw $2,%%mm7\n\t"
659
    "movq %%mm7,"OC_MEM_OFFS(0x18,buf)"\n\t"
660
    "psraw $2,%%mm4\n\t"
661
    "movq %%mm4,"OC_MEM_OFFS(0x38,buf)"\n\t"
662
    /*Final transpose and zig-zag.*/
663
#define OC_ZZ_LOAD_ROW_LO(_row,_reg) \
664
    "movq "OC_MEM_OFFS(16*_row,buf)","_reg"\n\t" \
665

666
#define OC_ZZ_LOAD_ROW_HI(_row,_reg) \
667
    "movq "OC_MEM_OFFS(16*_row+8,buf)","_reg"\n\t" \
668

669
    OC_TRANSPOSE_ZIG_ZAG_MMXEXT
670
#undef OC_ZZ_LOAD_ROW_LO
671
#undef OC_ZZ_LOAD_ROW_HI
672
    :[a]"=&r"(a),[buf]"=m"(OC_ARRAY_OPERAND(ogg_int16_t,buf,64))
673
    :[y]"r"(_y),[x]"r"(_x)
674
    :"memory"
675
  );
676
}
677

678
#endif
679

680
Product

Resources

Company
