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) 2002-2009 *
9 * by the Xiph.Org Foundation and contributors http://www.xiph.org/ *
10 * *
11 ********************************************************************
12
13 function:
14 last mod: Id
15
16 ********************************************************************/
17
18 /*MMX acceleration of Theora's iDCT.
19 Originally written by Rudolf Marek, based on code from On2's VP3.*/
20 #include "x86int.h"
21 #include "../dct.h"
22
23 #if defined(OC_X86_ASM)
24
25 /*These are offsets into the table of constants below.*/
26 /*7 rows of cosines, in order: pi/16 * (1 ... 7).*/
27 #define OC_COSINE_OFFSET (0)
28 /*A row of 8's.*/
29 #define OC_EIGHT_OFFSET (56)
30
31
32
33 /*A table of constants used by the MMX routines.*/
34 static const ogg_uint16_t __attribute__((aligned(8),used))
35 OC_IDCT_CONSTS[(7+1)*4]={
36 (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7,
37 (ogg_uint16_t)OC_C1S7,(ogg_uint16_t)OC_C1S7,
38 (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6,
39 (ogg_uint16_t)OC_C2S6,(ogg_uint16_t)OC_C2S6,
40 (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5,
41 (ogg_uint16_t)OC_C3S5,(ogg_uint16_t)OC_C3S5,
42 (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4,
43 (ogg_uint16_t)OC_C4S4,(ogg_uint16_t)OC_C4S4,
44 (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3,
45 (ogg_uint16_t)OC_C5S3,(ogg_uint16_t)OC_C5S3,
46 (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2,
47 (ogg_uint16_t)OC_C6S2,(ogg_uint16_t)OC_C6S2,
48 (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1,
49 (ogg_uint16_t)OC_C7S1,(ogg_uint16_t)OC_C7S1,
50 8, 8, 8, 8
51 };
52
53 /*Converts the expression in the argument to a string.*/
54 #define OC_M2STR(_s) #_s
55
56 /*38 cycles*/
57 #define OC_IDCT_BEGIN \
58 "#OC_IDCT_BEGIN\n\t" \
59 "movq "OC_I(3)",%%mm2\n\t" \
60 "movq "OC_C(3)",%%mm6\n\t" \
61 "movq %%mm2,%%mm4\n\t" \
62 "movq "OC_J(5)",%%mm7\n\t" \
63 "pmulhw %%mm6,%%mm4\n\t" \
64 "movq "OC_C(5)",%%mm1\n\t" \
65 "pmulhw %%mm7,%%mm6\n\t" \
66 "movq %%mm1,%%mm5\n\t" \
67 "pmulhw %%mm2,%%mm1\n\t" \
68 "movq "OC_I(1)",%%mm3\n\t" \
69 "pmulhw %%mm7,%%mm5\n\t" \
70 "movq "OC_C(1)",%%mm0\n\t" \
71 "paddw %%mm2,%%mm4\n\t" \
72 "paddw %%mm7,%%mm6\n\t" \
73 "paddw %%mm1,%%mm2\n\t" \
74 "movq "OC_J(7)",%%mm1\n\t" \
75 "paddw %%mm5,%%mm7\n\t" \
76 "movq %%mm0,%%mm5\n\t" \
77 "pmulhw %%mm3,%%mm0\n\t" \
78 "paddw %%mm7,%%mm4\n\t" \
79 "pmulhw %%mm1,%%mm5\n\t" \
80 "movq "OC_C(7)",%%mm7\n\t" \
81 "psubw %%mm2,%%mm6\n\t" \
82 "paddw %%mm3,%%mm0\n\t" \
83 "pmulhw %%mm7,%%mm3\n\t" \
84 "movq "OC_I(2)",%%mm2\n\t" \
85 "pmulhw %%mm1,%%mm7\n\t" \
86 "paddw %%mm1,%%mm5\n\t" \
87 "movq %%mm2,%%mm1\n\t" \
88 "pmulhw "OC_C(2)",%%mm2\n\t" \
89 "psubw %%mm5,%%mm3\n\t" \
90 "movq "OC_J(6)",%%mm5\n\t" \
91 "paddw %%mm7,%%mm0\n\t" \
92 "movq %%mm5,%%mm7\n\t" \
93 "psubw %%mm4,%%mm0\n\t" \
94 "pmulhw "OC_C(2)",%%mm5\n\t" \
95 "paddw %%mm1,%%mm2\n\t" \
96 "pmulhw "OC_C(6)",%%mm1\n\t" \
97 "paddw %%mm4,%%mm4\n\t" \
98 "paddw %%mm0,%%mm4\n\t" \
99 "psubw %%mm6,%%mm3\n\t" \
100 "paddw %%mm7,%%mm5\n\t" \
101 "paddw %%mm6,%%mm6\n\t" \
102 "pmulhw "OC_C(6)",%%mm7\n\t" \
103 "paddw %%mm3,%%mm6\n\t" \
104 "movq %%mm4,"OC_I(1)"\n\t" \
105 "psubw %%mm5,%%mm1\n\t" \
106 "movq "OC_C(4)",%%mm4\n\t" \
107 "movq %%mm3,%%mm5\n\t" \
108 "pmulhw %%mm4,%%mm3\n\t" \
109 "paddw %%mm2,%%mm7\n\t" \
110 "movq %%mm6,"OC_I(2)"\n\t" \
111 "movq %%mm0,%%mm2\n\t" \
112 "movq "OC_I(0)",%%mm6\n\t" \
113 "pmulhw %%mm4,%%mm0\n\t" \
114 "paddw %%mm3,%%mm5\n\t" \
115 "movq "OC_J(4)",%%mm3\n\t" \
116 "psubw %%mm1,%%mm5\n\t" \
117 "paddw %%mm0,%%mm2\n\t" \
118 "psubw %%mm3,%%mm6\n\t" \
119 "movq %%mm6,%%mm0\n\t" \
120 "pmulhw %%mm4,%%mm6\n\t" \
121 "paddw %%mm3,%%mm3\n\t" \
122 "paddw %%mm1,%%mm1\n\t" \
123 "paddw %%mm0,%%mm3\n\t" \
124 "paddw %%mm5,%%mm1\n\t" \
125 "pmulhw %%mm3,%%mm4\n\t" \
126 "paddw %%mm0,%%mm6\n\t" \
127 "psubw %%mm2,%%mm6\n\t" \
128 "paddw %%mm2,%%mm2\n\t" \
129 "movq "OC_I(1)",%%mm0\n\t" \
130 "paddw %%mm6,%%mm2\n\t" \
131 "paddw %%mm3,%%mm4\n\t" \
132 "psubw %%mm1,%%mm2\n\t" \
133 "#end OC_IDCT_BEGIN\n\t" \
134
135 /*38+8=46 cycles.*/
136 #define OC_ROW_IDCT \
137 "#OC_ROW_IDCT\n" \
138 OC_IDCT_BEGIN \
139 /*r3=D'*/ \
140 "movq "OC_I(2)",%%mm3\n\t" \
141 /*r4=E'=E-G*/ \
142 "psubw %%mm7,%%mm4\n\t" \
143 /*r1=H'+H'*/ \
144 "paddw %%mm1,%%mm1\n\t" \
145 /*r7=G+G*/ \
146 "paddw %%mm7,%%mm7\n\t" \
147 /*r1=R1=A''+H'*/ \
148 "paddw %%mm2,%%mm1\n\t" \
149 /*r7=G'=E+G*/ \
150 "paddw %%mm4,%%mm7\n\t" \
151 /*r4=R4=E'-D'*/ \
152 "psubw %%mm3,%%mm4\n\t" \
153 "paddw %%mm3,%%mm3\n\t" \
154 /*r6=R6=F'-B''*/ \
155 "psubw %%mm5,%%mm6\n\t" \
156 "paddw %%mm5,%%mm5\n\t" \
157 /*r3=R3=E'+D'*/ \
158 "paddw %%mm4,%%mm3\n\t" \
159 /*r5=R5=F'+B''*/ \
160 "paddw %%mm6,%%mm5\n\t" \
161 /*r7=R7=G'-C'*/ \
162 "psubw %%mm0,%%mm7\n\t" \
163 "paddw %%mm0,%%mm0\n\t" \
164 /*Save R1.*/ \
165 "movq %%mm1,"OC_I(1)"\n\t" \
166 /*r0=R0=G.+C.*/ \
167 "paddw %%mm7,%%mm0\n\t" \
168 "#end OC_ROW_IDCT\n\t" \
169
170 /*The following macro does two 4x4 transposes in place.
171 At entry, we assume:
172 r0 = a3 a2 a1 a0
173 I(1) = b3 b2 b1 b0
174 r2 = c3 c2 c1 c0
175 r3 = d3 d2 d1 d0
176
177 r4 = e3 e2 e1 e0
178 r5 = f3 f2 f1 f0
179 r6 = g3 g2 g1 g0
180 r7 = h3 h2 h1 h0
181
182 At exit, we have:
183 I(0) = d0 c0 b0 a0
184 I(1) = d1 c1 b1 a1
185 I(2) = d2 c2 b2 a2
186 I(3) = d3 c3 b3 a3
187
188 J(4) = h0 g0 f0 e0
189 J(5) = h1 g1 f1 e1
190 J(6) = h2 g2 f2 e2
191 J(7) = h3 g3 f3 e3
192
193 I(0) I(1) I(2) I(3) is the transpose of r0 I(1) r2 r3.
194 J(4) J(5) J(6) J(7) is the transpose of r4 r5 r6 r7.
195
196 Since r1 is free at entry, we calculate the Js first.*/
197 /*19 cycles.*/
198 #define OC_TRANSPOSE \
199 "#OC_TRANSPOSE\n\t" \
200 "movq %%mm4,%%mm1\n\t" \
201 "punpcklwd %%mm5,%%mm4\n\t" \
202 "movq %%mm0,"OC_I(0)"\n\t" \
203 "punpckhwd %%mm5,%%mm1\n\t" \
204 "movq %%mm6,%%mm0\n\t" \
205 "punpcklwd %%mm7,%%mm6\n\t" \
206 "movq %%mm4,%%mm5\n\t" \
207 "punpckldq %%mm6,%%mm4\n\t" \
208 "punpckhdq %%mm6,%%mm5\n\t" \
209 "movq %%mm1,%%mm6\n\t" \
210 "movq %%mm4,"OC_J(4)"\n\t" \
211 "punpckhwd %%mm7,%%mm0\n\t" \
212 "movq %%mm5,"OC_J(5)"\n\t" \
213 "punpckhdq %%mm0,%%mm6\n\t" \
214 "movq "OC_I(0)",%%mm4\n\t" \
215 "punpckldq %%mm0,%%mm1\n\t" \
216 "movq "OC_I(1)",%%mm5\n\t" \
217 "movq %%mm4,%%mm0\n\t" \
218 "movq %%mm6,"OC_J(7)"\n\t" \
219 "punpcklwd %%mm5,%%mm0\n\t" \
220 "movq %%mm1,"OC_J(6)"\n\t" \
221 "punpckhwd %%mm5,%%mm4\n\t" \
222 "movq %%mm2,%%mm5\n\t" \
223 "punpcklwd %%mm3,%%mm2\n\t" \
224 "movq %%mm0,%%mm1\n\t" \
225 "punpckldq %%mm2,%%mm0\n\t" \
226 "punpckhdq %%mm2,%%mm1\n\t" \
227 "movq %%mm4,%%mm2\n\t" \
228 "movq %%mm0,"OC_I(0)"\n\t" \
229 "punpckhwd %%mm3,%%mm5\n\t" \
230 "movq %%mm1,"OC_I(1)"\n\t" \
231 "punpckhdq %%mm5,%%mm4\n\t" \
232 "punpckldq %%mm5,%%mm2\n\t" \
233 "movq %%mm4,"OC_I(3)"\n\t" \
234 "movq %%mm2,"OC_I(2)"\n\t" \
235 "#end OC_TRANSPOSE\n\t" \
236
237 /*38+19=57 cycles.*/
238 #define OC_COLUMN_IDCT \
239 "#OC_COLUMN_IDCT\n" \
240 OC_IDCT_BEGIN \
241 "paddw "OC_8",%%mm2\n\t" \
242 /*r1=H'+H'*/ \
243 "paddw %%mm1,%%mm1\n\t" \
244 /*r1=R1=A''+H'*/ \
245 "paddw %%mm2,%%mm1\n\t" \
246 /*r2=NR2*/ \
247 "psraw $4,%%mm2\n\t" \
248 /*r4=E'=E-G*/ \
249 "psubw %%mm7,%%mm4\n\t" \
250 /*r1=NR1*/ \
251 "psraw $4,%%mm1\n\t" \
252 /*r3=D'*/ \
253 "movq "OC_I(2)",%%mm3\n\t" \
254 /*r7=G+G*/ \
255 "paddw %%mm7,%%mm7\n\t" \
256 /*Store NR2 at I(2).*/ \
257 "movq %%mm2,"OC_I(2)"\n\t" \
258 /*r7=G'=E+G*/ \
259 "paddw %%mm4,%%mm7\n\t" \
260 /*Store NR1 at I(1).*/ \
261 "movq %%mm1,"OC_I(1)"\n\t" \
262 /*r4=R4=E'-D'*/ \
263 "psubw %%mm3,%%mm4\n\t" \
264 "paddw "OC_8",%%mm4\n\t" \
265 /*r3=D'+D'*/ \
266 "paddw %%mm3,%%mm3\n\t" \
267 /*r3=R3=E'+D'*/ \
268 "paddw %%mm4,%%mm3\n\t" \
269 /*r4=NR4*/ \
270 "psraw $4,%%mm4\n\t" \
271 /*r6=R6=F'-B''*/ \
272 "psubw %%mm5,%%mm6\n\t" \
273 /*r3=NR3*/ \
274 "psraw $4,%%mm3\n\t" \
275 "paddw "OC_8",%%mm6\n\t" \
276 /*r5=B''+B''*/ \
277 "paddw %%mm5,%%mm5\n\t" \
278 /*r5=R5=F'+B''*/ \
279 "paddw %%mm6,%%mm5\n\t" \
280 /*r6=NR6*/ \
281 "psraw $4,%%mm6\n\t" \
282 /*Store NR4 at J(4).*/ \
283 "movq %%mm4,"OC_J(4)"\n\t" \
284 /*r5=NR5*/ \
285 "psraw $4,%%mm5\n\t" \
286 /*Store NR3 at I(3).*/ \
287 "movq %%mm3,"OC_I(3)"\n\t" \
288 /*r7=R7=G'-C'*/ \
289 "psubw %%mm0,%%mm7\n\t" \
290 "paddw "OC_8",%%mm7\n\t" \
291 /*r0=C'+C'*/ \
292 "paddw %%mm0,%%mm0\n\t" \
293 /*r0=R0=G'+C'*/ \
294 "paddw %%mm7,%%mm0\n\t" \
295 /*r7=NR7*/ \
296 "psraw $4,%%mm7\n\t" \
297 /*Store NR6 at J(6).*/ \
298 "movq %%mm6,"OC_J(6)"\n\t" \
299 /*r0=NR0*/ \
300 "psraw $4,%%mm0\n\t" \
301 /*Store NR5 at J(5).*/ \
302 "movq %%mm5,"OC_J(5)"\n\t" \
303 /*Store NR7 at J(7).*/ \
304 "movq %%mm7,"OC_J(7)"\n\t" \
305 /*Store NR0 at I(0).*/ \
306 "movq %%mm0,"OC_I(0)"\n\t" \
307 "#end OC_COLUMN_IDCT\n\t" \
308
309 #define OC_MID(_m,_i) OC_M2STR(_m+(_i)*8)"(%[c])"
310 #define OC_C(_i) OC_MID(OC_COSINE_OFFSET,_i-1)
311 #define OC_8 OC_MID(OC_EIGHT_OFFSET,0)
312
oc_idct8x8_slow(ogg_int16_t _y[64])313 static void oc_idct8x8_slow(ogg_int16_t _y[64]){
314 /*This routine accepts an 8x8 matrix, but in partially transposed form.
315 Every 4x4 block is transposed.*/
316 __asm__ __volatile__(
317 #define OC_I(_k) OC_M2STR((_k*16))"(%[y])"
318 #define OC_J(_k) OC_M2STR(((_k-4)*16)+8)"(%[y])"
319 OC_ROW_IDCT
320 OC_TRANSPOSE
321 #undef OC_I
322 #undef OC_J
323 #define OC_I(_k) OC_M2STR((_k*16)+64)"(%[y])"
324 #define OC_J(_k) OC_M2STR(((_k-4)*16)+72)"(%[y])"
325 OC_ROW_IDCT
326 OC_TRANSPOSE
327 #undef OC_I
328 #undef OC_J
329 #define OC_I(_k) OC_M2STR((_k*16))"(%[y])"
330 #define OC_J(_k) OC_I(_k)
331 OC_COLUMN_IDCT
332 #undef OC_I
333 #undef OC_J
334 #define OC_I(_k) OC_M2STR((_k*16)+8)"(%[y])"
335 #define OC_J(_k) OC_I(_k)
336 OC_COLUMN_IDCT
337 #undef OC_I
338 #undef OC_J
339 :
340 :[y]"r"(_y),[c]"r"(OC_IDCT_CONSTS)
341 );
342 }
343
344 /*25 cycles.*/
345 #define OC_IDCT_BEGIN_10 \
346 "#OC_IDCT_BEGIN_10\n\t" \
347 "movq "OC_I(3)",%%mm2\n\t" \
348 "nop\n\t" \
349 "movq "OC_C(3)",%%mm6\n\t" \
350 "movq %%mm2,%%mm4\n\t" \
351 "movq "OC_C(5)",%%mm1\n\t" \
352 "pmulhw %%mm6,%%mm4\n\t" \
353 "movq "OC_I(1)",%%mm3\n\t" \
354 "pmulhw %%mm2,%%mm1\n\t" \
355 "movq "OC_C(1)",%%mm0\n\t" \
356 "paddw %%mm2,%%mm4\n\t" \
357 "pxor %%mm6,%%mm6\n\t" \
358 "paddw %%mm1,%%mm2\n\t" \
359 "movq "OC_I(2)",%%mm5\n\t" \
360 "pmulhw %%mm3,%%mm0\n\t" \
361 "movq %%mm5,%%mm1\n\t" \
362 "paddw %%mm3,%%mm0\n\t" \
363 "pmulhw "OC_C(7)",%%mm3\n\t" \
364 "psubw %%mm2,%%mm6\n\t" \
365 "pmulhw "OC_C(2)",%%mm5\n\t" \
366 "psubw %%mm4,%%mm0\n\t" \
367 "movq "OC_I(2)",%%mm7\n\t" \
368 "paddw %%mm4,%%mm4\n\t" \
369 "paddw %%mm5,%%mm7\n\t" \
370 "paddw %%mm0,%%mm4\n\t" \
371 "pmulhw "OC_C(6)",%%mm1\n\t" \
372 "psubw %%mm6,%%mm3\n\t" \
373 "movq %%mm4,"OC_I(1)"\n\t" \
374 "paddw %%mm6,%%mm6\n\t" \
375 "movq "OC_C(4)",%%mm4\n\t" \
376 "paddw %%mm3,%%mm6\n\t" \
377 "movq %%mm3,%%mm5\n\t" \
378 "pmulhw %%mm4,%%mm3\n\t" \
379 "movq %%mm6,"OC_I(2)"\n\t" \
380 "movq %%mm0,%%mm2\n\t" \
381 "movq "OC_I(0)",%%mm6\n\t" \
382 "pmulhw %%mm4,%%mm0\n\t" \
383 "paddw %%mm3,%%mm5\n\t" \
384 "paddw %%mm0,%%mm2\n\t" \
385 "psubw %%mm1,%%mm5\n\t" \
386 "pmulhw %%mm4,%%mm6\n\t" \
387 "paddw "OC_I(0)",%%mm6\n\t" \
388 "paddw %%mm1,%%mm1\n\t" \
389 "movq %%mm6,%%mm4\n\t" \
390 "paddw %%mm5,%%mm1\n\t" \
391 "psubw %%mm2,%%mm6\n\t" \
392 "paddw %%mm2,%%mm2\n\t" \
393 "movq "OC_I(1)",%%mm0\n\t" \
394 "paddw %%mm6,%%mm2\n\t" \
395 "psubw %%mm1,%%mm2\n\t" \
396 "nop\n\t" \
397 "#end OC_IDCT_BEGIN_10\n\t" \
398
399 /*25+8=33 cycles.*/
400 #define OC_ROW_IDCT_10 \
401 "#OC_ROW_IDCT_10\n\t" \
402 OC_IDCT_BEGIN_10 \
403 /*r3=D'*/ \
404 "movq "OC_I(2)",%%mm3\n\t" \
405 /*r4=E'=E-G*/ \
406 "psubw %%mm7,%%mm4\n\t" \
407 /*r1=H'+H'*/ \
408 "paddw %%mm1,%%mm1\n\t" \
409 /*r7=G+G*/ \
410 "paddw %%mm7,%%mm7\n\t" \
411 /*r1=R1=A''+H'*/ \
412 "paddw %%mm2,%%mm1\n\t" \
413 /*r7=G'=E+G*/ \
414 "paddw %%mm4,%%mm7\n\t" \
415 /*r4=R4=E'-D'*/ \
416 "psubw %%mm3,%%mm4\n\t" \
417 "paddw %%mm3,%%mm3\n\t" \
418 /*r6=R6=F'-B''*/ \
419 "psubw %%mm5,%%mm6\n\t" \
420 "paddw %%mm5,%%mm5\n\t" \
421 /*r3=R3=E'+D'*/ \
422 "paddw %%mm4,%%mm3\n\t" \
423 /*r5=R5=F'+B''*/ \
424 "paddw %%mm6,%%mm5\n\t" \
425 /*r7=R7=G'-C'*/ \
426 "psubw %%mm0,%%mm7\n\t" \
427 "paddw %%mm0,%%mm0\n\t" \
428 /*Save R1.*/ \
429 "movq %%mm1,"OC_I(1)"\n\t" \
430 /*r0=R0=G'+C'*/ \
431 "paddw %%mm7,%%mm0\n\t" \
432 "#end OC_ROW_IDCT_10\n\t" \
433
434 /*25+19=44 cycles'*/
435 #define OC_COLUMN_IDCT_10 \
436 "#OC_COLUMN_IDCT_10\n\t" \
437 OC_IDCT_BEGIN_10 \
438 "paddw "OC_8",%%mm2\n\t" \
439 /*r1=H'+H'*/ \
440 "paddw %%mm1,%%mm1\n\t" \
441 /*r1=R1=A''+H'*/ \
442 "paddw %%mm2,%%mm1\n\t" \
443 /*r2=NR2*/ \
444 "psraw $4,%%mm2\n\t" \
445 /*r4=E'=E-G*/ \
446 "psubw %%mm7,%%mm4\n\t" \
447 /*r1=NR1*/ \
448 "psraw $4,%%mm1\n\t" \
449 /*r3=D'*/ \
450 "movq "OC_I(2)",%%mm3\n\t" \
451 /*r7=G+G*/ \
452 "paddw %%mm7,%%mm7\n\t" \
453 /*Store NR2 at I(2).*/ \
454 "movq %%mm2,"OC_I(2)"\n\t" \
455 /*r7=G'=E+G*/ \
456 "paddw %%mm4,%%mm7\n\t" \
457 /*Store NR1 at I(1).*/ \
458 "movq %%mm1,"OC_I(1)"\n\t" \
459 /*r4=R4=E'-D'*/ \
460 "psubw %%mm3,%%mm4\n\t" \
461 "paddw "OC_8",%%mm4\n\t" \
462 /*r3=D'+D'*/ \
463 "paddw %%mm3,%%mm3\n\t" \
464 /*r3=R3=E'+D'*/ \
465 "paddw %%mm4,%%mm3\n\t" \
466 /*r4=NR4*/ \
467 "psraw $4,%%mm4\n\t" \
468 /*r6=R6=F'-B''*/ \
469 "psubw %%mm5,%%mm6\n\t" \
470 /*r3=NR3*/ \
471 "psraw $4,%%mm3\n\t" \
472 "paddw "OC_8",%%mm6\n\t" \
473 /*r5=B''+B''*/ \
474 "paddw %%mm5,%%mm5\n\t" \
475 /*r5=R5=F'+B''*/ \
476 "paddw %%mm6,%%mm5\n\t" \
477 /*r6=NR6*/ \
478 "psraw $4,%%mm6\n\t" \
479 /*Store NR4 at J(4).*/ \
480 "movq %%mm4,"OC_J(4)"\n\t" \
481 /*r5=NR5*/ \
482 "psraw $4,%%mm5\n\t" \
483 /*Store NR3 at I(3).*/ \
484 "movq %%mm3,"OC_I(3)"\n\t" \
485 /*r7=R7=G'-C'*/ \
486 "psubw %%mm0,%%mm7\n\t" \
487 "paddw "OC_8",%%mm7\n\t" \
488 /*r0=C'+C'*/ \
489 "paddw %%mm0,%%mm0\n\t" \
490 /*r0=R0=G'+C'*/ \
491 "paddw %%mm7,%%mm0\n\t" \
492 /*r7=NR7*/ \
493 "psraw $4,%%mm7\n\t" \
494 /*Store NR6 at J(6).*/ \
495 "movq %%mm6,"OC_J(6)"\n\t" \
496 /*r0=NR0*/ \
497 "psraw $4,%%mm0\n\t" \
498 /*Store NR5 at J(5).*/ \
499 "movq %%mm5,"OC_J(5)"\n\t" \
500 /*Store NR7 at J(7).*/ \
501 "movq %%mm7,"OC_J(7)"\n\t" \
502 /*Store NR0 at I(0).*/ \
503 "movq %%mm0,"OC_I(0)"\n\t" \
504 "#end OC_COLUMN_IDCT_10\n\t" \
505
oc_idct8x8_10(ogg_int16_t _y[64])506 static void oc_idct8x8_10(ogg_int16_t _y[64]){
507 __asm__ __volatile__(
508 #define OC_I(_k) OC_M2STR((_k*16))"(%[y])"
509 #define OC_J(_k) OC_M2STR(((_k-4)*16)+8)"(%[y])"
510 /*Done with dequant, descramble, and partial transpose.
511 Now do the iDCT itself.*/
512 OC_ROW_IDCT_10
513 OC_TRANSPOSE
514 #undef OC_I
515 #undef OC_J
516 #define OC_I(_k) OC_M2STR((_k*16))"(%[y])"
517 #define OC_J(_k) OC_I(_k)
518 OC_COLUMN_IDCT_10
519 #undef OC_I
520 #undef OC_J
521 #define OC_I(_k) OC_M2STR((_k*16)+8)"(%[y])"
522 #define OC_J(_k) OC_I(_k)
523 OC_COLUMN_IDCT_10
524 #undef OC_I
525 #undef OC_J
526 :
527 :[y]"r"(_y),[c]"r"(OC_IDCT_CONSTS)
528 );
529 }
530
531 /*Performs an inverse 8x8 Type-II DCT transform.
532 The input is assumed to be scaled by a factor of 4 relative to orthonormal
533 version of the transform.*/
oc_idct8x8_mmx(ogg_int16_t _y[64],int _last_zzi)534 void oc_idct8x8_mmx(ogg_int16_t _y[64],int _last_zzi){
535 /*_last_zzi is subtly different from an actual count of the number of
536 coefficients we decoded for this block.
537 It contains the value of zzi BEFORE the final token in the block was
538 decoded.
539 In most cases this is an EOB token (the continuation of an EOB run from a
540 previous block counts), and so this is the same as the coefficient count.
541 However, in the case that the last token was NOT an EOB token, but filled
542 the block up with exactly 64 coefficients, _last_zzi will be less than 64.
543 Provided the last token was not a pure zero run, the minimum value it can
544 be is 46, and so that doesn't affect any of the cases in this routine.
545 However, if the last token WAS a pure zero run of length 63, then _last_zzi
546 will be 1 while the number of coefficients decoded is 64.
547 Thus, we will trigger the following special case, where the real
548 coefficient count would not.
549 Note also that a zero run of length 64 will give _last_zzi a value of 0,
550 but we still process the DC coefficient, which might have a non-zero value
551 due to DC prediction.
552 Although convoluted, this is arguably the correct behavior: it allows us to
553 use a smaller transform when the block ends with a long zero run instead
554 of a normal EOB token.
555 It could be smarter... multiple separate zero runs at the end of a block
556 will fool it, but an encoder that generates these really deserves what it
557 gets.
558 Needless to say we inherited this approach from VP3.*/
559 /*Then perform the iDCT.*/
560 if(_last_zzi<10)oc_idct8x8_10(_y);
561 else oc_idct8x8_slow(_y);
562 }
563
564 #endif
565