1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
|
#if !defined(_x86_mmxloop_H)
# define _x86_mmxloop_H (1)
# include <stddef.h>
# include "x86int.h"
#if defined(OC_X86_ASM)
/*On entry, mm0={a0,...,a7}, mm1={b0,...,b7}, mm2={c0,...,c7}, mm3={d0,...d7}.
On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}; mm0 and mm3 are clobbered.*/
#define OC_LOOP_FILTER8_MMX \
"#OC_LOOP_FILTER8_MMX\n\t" \
/*mm7=0*/ \
"pxor %%mm7,%%mm7\n\t" \
/*mm6:mm0={a0,...,a7}*/ \
"movq %%mm0,%%mm6\n\t" \
"punpcklbw %%mm7,%%mm0\n\t" \
"punpckhbw %%mm7,%%mm6\n\t" \
/*mm3:mm5={d0,...,d7}*/ \
"movq %%mm3,%%mm5\n\t" \
"punpcklbw %%mm7,%%mm3\n\t" \
"punpckhbw %%mm7,%%mm5\n\t" \
/*mm6:mm0={a0-d0,...,a7-d7}*/ \
"psubw %%mm3,%%mm0\n\t" \
"psubw %%mm5,%%mm6\n\t" \
/*mm3:mm1={b0,...,b7}*/ \
"movq %%mm1,%%mm3\n\t" \
"punpcklbw %%mm7,%%mm1\n\t" \
"movq %%mm2,%%mm4\n\t" \
"punpckhbw %%mm7,%%mm3\n\t" \
/*mm5:mm4={c0,...,c7}*/ \
"movq %%mm2,%%mm5\n\t" \
"punpcklbw %%mm7,%%mm4\n\t" \
"punpckhbw %%mm7,%%mm5\n\t" \
/*mm7={3}x4 \
mm5:mm4={c0-b0,...,c7-b7}*/ \
"pcmpeqw %%mm7,%%mm7\n\t" \
"psubw %%mm1,%%mm4\n\t" \
"psrlw $14,%%mm7\n\t" \
"psubw %%mm3,%%mm5\n\t" \
/*Scale by 3.*/ \
"pmullw %%mm7,%%mm4\n\t" \
"pmullw %%mm7,%%mm5\n\t" \
/*mm7={4}x4 \
mm5:mm4=f={a0-d0+3*(c0-b0),...,a7-d7+3*(c7-b7)}*/ \
"psrlw $1,%%mm7\n\t" \
"paddw %%mm0,%%mm4\n\t" \
"psllw $2,%%mm7\n\t" \
"movq (%[ll]),%%mm0\n\t" \
"paddw %%mm6,%%mm5\n\t" \
/*R_i has the range [-127,128], so we compute -R_i instead. \
mm4=-R_i=-(f+4>>3)=0xFF^(f-4>>3)*/ \
"psubw %%mm7,%%mm4\n\t" \
"psubw %%mm7,%%mm5\n\t" \
"psraw $3,%%mm4\n\t" \
"psraw $3,%%mm5\n\t" \
"pcmpeqb %%mm7,%%mm7\n\t" \
"packsswb %%mm5,%%mm4\n\t" \
"pxor %%mm6,%%mm6\n\t" \
"pxor %%mm7,%%mm4\n\t" \
"packuswb %%mm3,%%mm1\n\t" \
/*Now compute lflim of -mm4 cf. Section 7.10 of the sepc.*/ \
/*There's no unsigned byte+signed byte with unsigned saturation op code, so \
we have to split things by sign (the other option is to work in 16 bits, \
but working in 8 bits gives much better parallelism). \
We compute abs(R_i), but save a mask of which terms were negative in mm6. \
Then we compute mm4=abs(lflim(R_i,L))=min(abs(R_i),max(2*L-abs(R_i),0)). \
Finally, we split mm4 into positive and negative pieces using the mask in \
mm6, and add and subtract them as appropriate.*/ \
/*mm4=abs(-R_i)*/ \
/*mm7=255-2*L*/ \
"pcmpgtb %%mm4,%%mm6\n\t" \
"psubb %%mm0,%%mm7\n\t" \
"pxor %%mm6,%%mm4\n\t" \
"psubb %%mm0,%%mm7\n\t" \
"psubb %%mm6,%%mm4\n\t" \
/*mm7=255-max(2*L-abs(R_i),0)*/ \
"paddusb %%mm4,%%mm7\n\t" \
/*mm4=min(abs(R_i),max(2*L-abs(R_i),0))*/ \
"paddusb %%mm7,%%mm4\n\t" \
"psubusb %%mm7,%%mm4\n\t" \
/*Now split mm4 by the original sign of -R_i.*/ \
"movq %%mm4,%%mm5\n\t" \
"pand %%mm6,%%mm4\n\t" \
"pandn %%mm5,%%mm6\n\t" \
/*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
/*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
"paddusb %%mm4,%%mm1\n\t" \
"psubusb %%mm4,%%mm2\n\t" \
"psubusb %%mm6,%%mm1\n\t" \
"paddusb %%mm6,%%mm2\n\t" \
/*On entry, mm0={a0,...,a7}, mm1={b0,...,b7}, mm2={c0,...,c7}, mm3={d0,...d7}.
On exit, mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)} and
mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}.
All other MMX registers are clobbered.*/
#define OC_LOOP_FILTER8_MMXEXT \
"#OC_LOOP_FILTER8_MMXEXT\n\t" \
/*R_i=(a_i-3*b_i+3*c_i-d_i+4>>3) has the range [-127,128], so we compute \
-R_i=(-a_i+3*b_i-3*c_i+d_i+3>>3) instead.*/ \
/*This first part is based on the transformation \
f = -(3*(c-b)+a-d+4>>3) \
= -(3*(c+255-b)+(a+255-d)+4-1020>>3) \
= -(3*(c+~b)+(a+~d)-1016>>3) \
= 127-(3*(c+~b)+(a+~d)>>3) \
= 128+~(3*(c+~b)+(a+~d)>>3) (mod 256). \
Although pavgb(a,b) = (a+b+1>>1) (biased up), we rely heavily on the \
fact that ~pavgb(~a,~b) = (a+b>>1) (biased down). \
Using this, the last expression above can be computed in 8 bits of working \
precision via: \
u = ~pavgb(~b,c); \
v = pavgb(b,~c); \
This mask is 0 or 0xFF, and controls whether t is biased up or down: \
m = u-v; \
t = m^pavgb(m^~a,m^d); \
f = 128+pavgb(pavgb(t,u),v); \
This required some careful analysis to ensure that carries are propagated \
correctly in all cases, but has been checked exhaustively.*/ \
/*input (a, b, c, d, ., ., ., .)*/ \
/*ff=0xFF; \
u=b; \
v=c; \
ll=255-2*L;*/ \
"pcmpeqb %%mm7,%%mm7\n\t" \
"movq %%mm1,%%mm4\n\t" \
"movq %%mm2,%%mm5\n\t" \
"movq (%[ll]),%%mm6\n\t" \
/*allocated u, v, ll, ff: (a, b, c, d, u, v, ll, ff)*/ \
/*u^=ff; \
v^=ff;*/ \
"pxor %%mm7,%%mm4\n\t" \
"pxor %%mm7,%%mm5\n\t" \
/*allocated ll: (a, b, c, d, u, v, ll, ff)*/ \
/*u=pavgb(u,c); \
v=pavgb(v,b);*/ \
"pavgb %%mm2,%%mm4\n\t" \
"pavgb %%mm1,%%mm5\n\t" \
/*u^=ff; \
a^=ff;*/ \
"pxor %%mm7,%%mm4\n\t" \
"pxor %%mm7,%%mm0\n\t" \
/*m=u-v;*/ \
"psubb %%mm5,%%mm4\n\t" \
/*freed u, allocated m: (a, b, c, d, m, v, ll, ff)*/ \
/*a^=m; \
d^=m;*/ \
"pxor %%mm4,%%mm0\n\t" \
"pxor %%mm4,%%mm3\n\t" \
/*t=pavgb(a,d);*/ \
"pavgb %%mm3,%%mm0\n\t" \
"psllw $7,%%mm7\n\t" \
/*freed a, d, ff, allocated t, of: (t, b, c, ., m, v, ll, of)*/ \
/*t^=m; \
u=m+v;*/ \
"pxor %%mm4,%%mm0\n\t" \
"paddb %%mm5,%%mm4\n\t" \
/*freed t, m, allocated f, u: (f, b, c, ., u, v, ll, of)*/ \
/*f=pavgb(f,u); \
of=128;*/ \
"pavgb %%mm4,%%mm0\n\t" \
"packsswb %%mm7,%%mm7\n\t" \
/*freed u, ff, allocated ll: (f, b, c, ., ll, v, ll, of)*/ \
/*f=pavgb(f,v);*/ \
"pavgb %%mm5,%%mm0\n\t" \
"movq %%mm7,%%mm3\n\t" \
"movq %%mm6,%%mm4\n\t" \
/*freed v, allocated of: (f, b, c, of, ll, ., ll, of)*/ \
/*Now compute lflim of R_i=-(128+mm0) cf. Section 7.10 of the sepc.*/ \
/*There's no unsigned byte+signed byte with unsigned saturation op code, so \
we have to split things by sign (the other option is to work in 16 bits, \
but staying in 8 bits gives much better parallelism).*/ \
/*Instead of adding the offset of 128 in mm3, we use it to split mm0. \
This is the same number of instructions as computing a mask and splitting \
after the lflim computation, but has shorter dependency chains.*/ \
/*mm0=R_i<0?-R_i:0 (denoted abs(R_i<0))\
mm3=R_i>0?R_i:0* (denoted abs(R_i>0))*/ \
"psubusb %%mm0,%%mm3\n\t" \
"psubusb %%mm7,%%mm0\n\t" \
/*mm6=255-max(2*L-abs(R_i<0),0) \
mm4=255-max(2*L-abs(R_i>0),0)*/ \
"paddusb %%mm3,%%mm4\n\t" \
"paddusb %%mm0,%%mm6\n\t" \
/*mm0=min(abs(R_i<0),max(2*L-abs(R_i<0),0)) \
mm3=min(abs(R_i>0),max(2*L-abs(R_i>0),0))*/ \
"paddusb %%mm4,%%mm3\n\t" \
"paddusb %%mm6,%%mm0\n\t" \
"psubusb %%mm4,%%mm3\n\t" \
"psubusb %%mm6,%%mm0\n\t" \
/*mm1={b0+lflim(R_0,L),...,b7+lflim(R_7,L)}*/ \
/*mm2={c0-lflim(R_0,L),...,c7-lflim(R_7,L)}*/ \
"paddusb %%mm3,%%mm1\n\t" \
"psubusb %%mm3,%%mm2\n\t" \
"psubusb %%mm0,%%mm1\n\t" \
"paddusb %%mm0,%%mm2\n\t" \
#define OC_LOOP_FILTER_V(_filter,_pix,_ystride,_ll) \
do{ \
ptrdiff_t ystride3__; \
__asm__ __volatile__( \
/*mm0={a0,...,a7}*/ \
"movq (%[pix]),%%mm0\n\t" \
/*ystride3=_ystride*3*/ \
"lea (%[ystride],%[ystride],2),%[ystride3]\n\t" \
/*mm3={d0,...,d7}*/ \
"movq (%[pix],%[ystride3]),%%mm3\n\t" \
/*mm1={b0,...,b7}*/ \
"movq (%[pix],%[ystride]),%%mm1\n\t" \
/*mm2={c0,...,c7}*/ \
"movq (%[pix],%[ystride],2),%%mm2\n\t" \
_filter \
/*Write it back out.*/ \
"movq %%mm1,(%[pix],%[ystride])\n\t" \
"movq %%mm2,(%[pix],%[ystride],2)\n\t" \
:[ystride3]"=&r"(ystride3__) \
:[pix]"r"(_pix-_ystride*2),[ystride]"r"((ptrdiff_t)(_ystride)), \
[ll]"r"(_ll) \
:"memory" \
); \
} \
while(0)
#define OC_LOOP_FILTER_H(_filter,_pix,_ystride,_ll) \
do{ \
unsigned char *pix__; \
ptrdiff_t ystride3__; \
ptrdiff_t d__; \
pix__=(_pix)-2; \
__asm__ __volatile__( \
/*x x x x d0 c0 b0 a0*/ \
"movd (%[pix]),%%mm0\n\t" \
/*x x x x d1 c1 b1 a1*/ \
"movd (%[pix],%[ystride]),%%mm1\n\t" \
/*ystride3=_ystride*3*/ \
"lea (%[ystride],%[ystride],2),%[ystride3]\n\t" \
/*x x x x d2 c2 b2 a2*/ \
"movd (%[pix],%[ystride],2),%%mm2\n\t" \
/*x x x x d3 c3 b3 a3*/ \
"lea (%[pix],%[ystride],4),%[d]\n\t" \
"movd (%[pix],%[ystride3]),%%mm3\n\t" \
/*x x x x d4 c4 b4 a4*/ \
"movd (%[d]),%%mm4\n\t" \
/*x x x x d5 c5 b5 a5*/ \
"movd (%[d],%[ystride]),%%mm5\n\t" \
/*x x x x d6 c6 b6 a6*/ \
"movd (%[d],%[ystride],2),%%mm6\n\t" \
/*x x x x d7 c7 b7 a7*/ \
"movd (%[d],%[ystride3]),%%mm7\n\t" \
/*mm0=d1 d0 c1 c0 b1 b0 a1 a0*/ \
"punpcklbw %%mm1,%%mm0\n\t" \
/*mm2=d3 d2 c3 c2 b3 b2 a3 a2*/ \
"punpcklbw %%mm3,%%mm2\n\t" \
/*mm3=d1 d0 c1 c0 b1 b0 a1 a0*/ \
"movq %%mm0,%%mm3\n\t" \
/*mm0=b3 b2 b1 b0 a3 a2 a1 a0*/ \
"punpcklwd %%mm2,%%mm0\n\t" \
/*mm3=d3 d2 d1 d0 c3 c2 c1 c0*/ \
"punpckhwd %%mm2,%%mm3\n\t" \
/*mm1=b3 b2 b1 b0 a3 a2 a1 a0*/ \
"movq %%mm0,%%mm1\n\t" \
/*mm4=d5 d4 c5 c4 b5 b4 a5 a4*/ \
"punpcklbw %%mm5,%%mm4\n\t" \
/*mm6=d7 d6 c7 c6 b7 b6 a7 a6*/ \
"punpcklbw %%mm7,%%mm6\n\t" \
/*mm5=d5 d4 c5 c4 b5 b4 a5 a4*/ \
"movq %%mm4,%%mm5\n\t" \
/*mm4=b7 b6 b5 b4 a7 a6 a5 a4*/ \
"punpcklwd %%mm6,%%mm4\n\t" \
/*mm5=d7 d6 d5 d4 c7 c6 c5 c4*/ \
"punpckhwd %%mm6,%%mm5\n\t" \
/*mm2=d3 d2 d1 d0 c3 c2 c1 c0*/ \
"movq %%mm3,%%mm2\n\t" \
/*mm0=a7 a6 a5 a4 a3 a2 a1 a0*/ \
"punpckldq %%mm4,%%mm0\n\t" \
/*mm1=b7 b6 b5 b4 b3 b2 b1 b0*/ \
"punpckhdq %%mm4,%%mm1\n\t" \
/*mm2=c7 c6 c5 c4 c3 c2 c1 c0*/ \
"punpckldq %%mm5,%%mm2\n\t" \
/*mm3=d7 d6 d5 d4 d3 d2 d1 d0*/ \
"punpckhdq %%mm5,%%mm3\n\t" \
_filter \
/*mm2={b0+R_0'',...,b7+R_7''}*/ \
"movq %%mm1,%%mm0\n\t" \
/*mm1={b0+R_0'',c0-R_0'',...,b3+R_3'',c3-R_3''}*/ \
"punpcklbw %%mm2,%%mm1\n\t" \
/*mm2={b4+R_4'',c4-R_4'',...,b7+R_7'',c7-R_7''}*/ \
"punpckhbw %%mm2,%%mm0\n\t" \
/*[d]=c1 b1 c0 b0*/ \
"movd %%mm1,%[d]\n\t" \
"movw %w[d],1(%[pix])\n\t" \
"psrlq $32,%%mm1\n\t" \
"shr $16,%[d]\n\t" \
"movw %w[d],1(%[pix],%[ystride])\n\t" \
/*[d]=c3 b3 c2 b2*/ \
"movd %%mm1,%[d]\n\t" \
"movw %w[d],1(%[pix],%[ystride],2)\n\t" \
"shr $16,%[d]\n\t" \
"movw %w[d],1(%[pix],%[ystride3])\n\t" \
"lea (%[pix],%[ystride],4),%[pix]\n\t" \
/*[d]=c5 b5 c4 b4*/ \
"movd %%mm0,%[d]\n\t" \
"movw %w[d],1(%[pix])\n\t" \
"psrlq $32,%%mm0\n\t" \
"shr $16,%[d]\n\t" \
"movw %w[d],1(%[pix],%[ystride])\n\t" \
/*[d]=c7 b7 c6 b6*/ \
"movd %%mm0,%[d]\n\t" \
"movw %w[d],1(%[pix],%[ystride],2)\n\t" \
"shr $16,%[d]\n\t" \
"movw %w[d],1(%[pix],%[ystride3])\n\t" \
:[pix]"+r"(pix__),[ystride3]"=&r"(ystride3__),[d]"=&r"(d__) \
:[ystride]"r"((ptrdiff_t)(_ystride)),[ll]"r"(_ll) \
:"memory" \
); \
} \
while(0)
# endif
#endif
|
