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
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
|
;*****************************************************************************
;* MMX optimized DSP utils
;*****************************************************************************
;* Copyright (c) 2000, 2001 Fabrice Bellard
;* Copyright (c) 2002-2004 Michael Niedermayer <michaelni@gmx.at>
;*
;* This file is part of FFmpeg.
;*
;* FFmpeg is free software; you can redistribute it and/or
;* modify it under the terms of the GNU Lesser General Public
;* License as published by the Free Software Foundation; either
;* version 2.1 of the License, or (at your option) any later version.
;*
;* FFmpeg is distributed in the hope that it will be useful,
;* but WITHOUT ANY WARRANTY; without even the implied warranty of
;* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
;* Lesser General Public License for more details.
;*
;* You should have received a copy of the GNU Lesser General Public
;* License along with FFmpeg; if not, write to the Free Software
;* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
;*****************************************************************************
%include "x86inc.asm"
%include "x86util.asm"
SECTION .text
%macro DIFF_PIXELS_1 4
movh %1, %3
movh %2, %4
punpcklbw %2, %1
punpcklbw %1, %1
psubw %1, %2
%endmacro
; %1=uint8_t *pix1, %2=uint8_t *pix2, %3=static offset, %4=stride, %5=stride*3
; %6=temporary storage location
; this macro requires $mmsize stack space (aligned) on %6 (except on SSE+x86-64)
%macro DIFF_PIXELS_8 6
DIFF_PIXELS_1 m0, m7, [%1 +%3], [%2 +%3]
DIFF_PIXELS_1 m1, m7, [%1+%4 +%3], [%2+%4 +%3]
DIFF_PIXELS_1 m2, m7, [%1+%4*2+%3], [%2+%4*2+%3]
add %1, %5
add %2, %5
DIFF_PIXELS_1 m3, m7, [%1 +%3], [%2 +%3]
DIFF_PIXELS_1 m4, m7, [%1+%4 +%3], [%2+%4 +%3]
DIFF_PIXELS_1 m5, m7, [%1+%4*2+%3], [%2+%4*2+%3]
DIFF_PIXELS_1 m6, m7, [%1+%5 +%3], [%2+%5 +%3]
%ifdef m8
DIFF_PIXELS_1 m7, m8, [%1+%4*4+%3], [%2+%4*4+%3]
%else
mova [%6], m0
DIFF_PIXELS_1 m7, m0, [%1+%4*4+%3], [%2+%4*4+%3]
mova m0, [%6]
%endif
sub %1, %5
sub %2, %5
%endmacro
%macro HADAMARD8 0
SUMSUB_BADC w, 0, 1, 2, 3
SUMSUB_BADC w, 4, 5, 6, 7
SUMSUB_BADC w, 0, 2, 1, 3
SUMSUB_BADC w, 4, 6, 5, 7
SUMSUB_BADC w, 0, 4, 1, 5
SUMSUB_BADC w, 2, 6, 3, 7
%endmacro
%macro ABS1_SUM 3
ABS1 %1, %2
paddusw %3, %1
%endmacro
%macro ABS2_SUM 6
ABS2 %1, %2, %3, %4
paddusw %5, %1
paddusw %6, %2
%endmacro
%macro ABS_SUM_8x8_64 1
ABS2 m0, m1, m8, m9
ABS2_SUM m2, m3, m8, m9, m0, m1
ABS2_SUM m4, m5, m8, m9, m0, m1
ABS2_SUM m6, m7, m8, m9, m0, m1
paddusw m0, m1
%endmacro
%macro ABS_SUM_8x8_32 1
mova [%1], m7
ABS1 m0, m7
ABS1 m1, m7
ABS1_SUM m2, m7, m0
ABS1_SUM m3, m7, m1
ABS1_SUM m4, m7, m0
ABS1_SUM m5, m7, m1
ABS1_SUM m6, m7, m0
mova m2, [%1]
ABS1_SUM m2, m7, m1
paddusw m0, m1
%endmacro
; FIXME: HSUM_* saturates at 64k, while an 8x8 hadamard or dct block can get up to
; about 100k on extreme inputs. But that's very unlikely to occur in natural video,
; and it's even more unlikely to not have any alternative mvs/modes with lower cost.
%macro HSUM_MMX 3
mova %2, %1
psrlq %1, 32
paddusw %1, %2
mova %2, %1
psrlq %1, 16
paddusw %1, %2
movd %3, %1
%endmacro
%macro HSUM_MMX2 3
pshufw %2, %1, 0xE
paddusw %1, %2
pshufw %2, %1, 0x1
paddusw %1, %2
movd %3, %1
%endmacro
%macro HSUM_SSE2 3
movhlps %2, %1
paddusw %1, %2
pshuflw %2, %1, 0xE
paddusw %1, %2
pshuflw %2, %1, 0x1
paddusw %1, %2
movd %3, %1
%endmacro
%macro STORE4 5
mova [%1+mmsize*0], %2
mova [%1+mmsize*1], %3
mova [%1+mmsize*2], %4
mova [%1+mmsize*3], %5
%endmacro
%macro LOAD4 5
mova %2, [%1+mmsize*0]
mova %3, [%1+mmsize*1]
mova %4, [%1+mmsize*2]
mova %5, [%1+mmsize*3]
%endmacro
%macro hadamard8_16_wrapper 3
cglobal hadamard8_diff_%1, 4, 4, %2
%ifndef m8
%assign pad %3*mmsize-(4+stack_offset&(mmsize-1))
SUB rsp, pad
%endif
call hadamard8x8_diff_%1
%ifndef m8
ADD rsp, pad
%endif
RET
cglobal hadamard8_diff16_%1, 5, 6, %2
%ifndef m8
%assign pad %3*mmsize-(4+stack_offset&(mmsize-1))
SUB rsp, pad
%endif
call hadamard8x8_diff_%1
mov r5d, eax
add r1, 8
add r2, 8
call hadamard8x8_diff_%1
add r5d, eax
cmp r4d, 16
jne .done
lea r1, [r1+r3*8-8]
lea r2, [r2+r3*8-8]
call hadamard8x8_diff_%1
add r5d, eax
add r1, 8
add r2, 8
call hadamard8x8_diff_%1
add r5d, eax
.done
mov eax, r5d
%ifndef m8
ADD rsp, pad
%endif
RET
%endmacro
%macro HADAMARD8_DIFF_MMX 1
ALIGN 16
; int hadamard8_diff_##cpu(void *s, uint8_t *src1, uint8_t *src2,
; int stride, int h)
; r0 = void *s = unused, int h = unused (always 8)
; note how r1, r2 and r3 are not clobbered in this function, so 16x16
; can simply call this 2x2x (and that's why we access rsp+gprsize
; everywhere, which is rsp of calling func
hadamard8x8_diff_%1:
lea r0, [r3*3]
; first 4x8 pixels
DIFF_PIXELS_8 r1, r2, 0, r3, r0, rsp+gprsize+0x60
HADAMARD8
mova [rsp+gprsize+0x60], m7
TRANSPOSE4x4W 0, 1, 2, 3, 7
STORE4 rsp+gprsize, m0, m1, m2, m3
mova m7, [rsp+gprsize+0x60]
TRANSPOSE4x4W 4, 5, 6, 7, 0
STORE4 rsp+gprsize+0x40, m4, m5, m6, m7
; second 4x8 pixels
DIFF_PIXELS_8 r1, r2, 4, r3, r0, rsp+gprsize+0x60
HADAMARD8
mova [rsp+gprsize+0x60], m7
TRANSPOSE4x4W 0, 1, 2, 3, 7
STORE4 rsp+gprsize+0x20, m0, m1, m2, m3
mova m7, [rsp+gprsize+0x60]
TRANSPOSE4x4W 4, 5, 6, 7, 0
LOAD4 rsp+gprsize+0x40, m0, m1, m2, m3
HADAMARD8
ABS_SUM_8x8_32 rsp+gprsize+0x60
mova [rsp+gprsize+0x60], m0
LOAD4 rsp+gprsize , m0, m1, m2, m3
LOAD4 rsp+gprsize+0x20, m4, m5, m6, m7
HADAMARD8
ABS_SUM_8x8_32 rsp+gprsize
paddusw m0, [rsp+gprsize+0x60]
HSUM m0, m1, eax
and rax, 0xFFFF
ret
hadamard8_16_wrapper %1, 0, 14
%endmacro
%macro HADAMARD8_DIFF_SSE2 2
hadamard8x8_diff_%1:
lea r0, [r3*3]
DIFF_PIXELS_8 r1, r2, 0, r3, r0, rsp+gprsize
HADAMARD8
%ifdef ARCH_X86_64
TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, 8
%else
TRANSPOSE8x8W 0, 1, 2, 3, 4, 5, 6, 7, [rsp+gprsize], [rsp+mmsize+gprsize]
%endif
HADAMARD8
ABS_SUM_8x8 rsp+gprsize
HSUM_SSE2 m0, m1, eax
and eax, 0xFFFF
ret
hadamard8_16_wrapper %1, %2, 3
%endmacro
INIT_MMX
%define ABS1 ABS1_MMX
%define HSUM HSUM_MMX
HADAMARD8_DIFF_MMX mmx
%define ABS1 ABS1_MMX2
%define HSUM HSUM_MMX2
HADAMARD8_DIFF_MMX mmx2
INIT_XMM
%define ABS2 ABS2_MMX2
%ifdef ARCH_X86_64
%define ABS_SUM_8x8 ABS_SUM_8x8_64
%else
%define ABS_SUM_8x8 ABS_SUM_8x8_32
%endif
HADAMARD8_DIFF_SSE2 sse2, 10
%define ABS2 ABS2_SSSE3
%define ABS_SUM_8x8 ABS_SUM_8x8_64
HADAMARD8_DIFF_SSE2 ssse3, 9
INIT_XMM
; sse16_sse2(void *v, uint8_t * pix1, uint8_t * pix2, int line_size, int h)
cglobal sse16_sse2, 5, 5, 8
shr r4d, 1
pxor m0, m0 ; mm0 = 0
pxor m7, m7 ; mm7 holds the sum
.next2lines ; FIXME why are these unaligned movs? pix1[] is aligned
movu m1, [r1 ] ; mm1 = pix1[0][0-15]
movu m2, [r2 ] ; mm2 = pix2[0][0-15]
movu m3, [r1+r3] ; mm3 = pix1[1][0-15]
movu m4, [r2+r3] ; mm4 = pix2[1][0-15]
; todo: mm1-mm2, mm3-mm4
; algo: subtract mm1 from mm2 with saturation and vice versa
; OR the result to get the absolute difference
mova m5, m1
mova m6, m3
psubusb m1, m2
psubusb m3, m4
psubusb m2, m5
psubusb m4, m6
por m2, m1
por m4, m3
; now convert to 16-bit vectors so we can square them
mova m1, m2
mova m3, m4
punpckhbw m2, m0
punpckhbw m4, m0
punpcklbw m1, m0 ; mm1 not spread over (mm1,mm2)
punpcklbw m3, m0 ; mm4 not spread over (mm3,mm4)
pmaddwd m2, m2
pmaddwd m4, m4
pmaddwd m1, m1
pmaddwd m3, m3
lea r1, [r1+r3*2] ; pix1 += 2*line_size
lea r2, [r2+r3*2] ; pix2 += 2*line_size
paddd m1, m2
paddd m3, m4
paddd m7, m1
paddd m7, m3
dec r4
jnz .next2lines
mova m1, m7
psrldq m7, 8 ; shift hi qword to lo
paddd m7, m1
mova m1, m7
psrldq m7, 4 ; shift hi dword to lo
paddd m7, m1
movd eax, m7 ; return value
RET
|
